This page lists 335 corrections that changed 1053 pages on this site. Every affected page carries the same entry in a Corrections block near the bottom, and the page data names it in a correction property. The log begins with the corrections of October 20 and 22, 2026; earlier fixes are not listed here.

Found something wrong? Use the correction form or write to manuel@aivortex.io. We check the claim against a primary source, change the page if it holds up, and add it here with the date.

AI in labor and employment law: the page said a January 2024 ruling made employers liable under Title VII; it now states what the court's orders show (age-discrimination claim, agency theory, 2024)

· 1 page

What changed

  • The page said a federal court ruled in January 2024 that an employer could be liable under Title VII for an AI vendor's screening tool, and that Workday could be treated as an employment agency or agent under Title VII, the ADEA and the ADA. We could not confirm the month or the Title VII holding from the court's own orders. The page now says the court allowed an age-discrimination (ADEA) disparate-impact claim to proceed against Workday on an agency theory in 2024, notes the May 16, 2025 preliminary collective certification, and says the case is still being litigated. (1 page)

Source

Mobley v. Workday, Inc., No. 3:23-cv-00770-RFL (N.D. Cal.), Dkt. 128, order granting preliminary collective certification of May 16, 2025, which summarizes the motion-to-dismiss order (Dkt. 80) and cites it as 740 F. Supp. 3d 796 (N.D. Cal. 2024) (read November 18, 2026)

Pages that carry this correction (1)

Georgia: State Bar Generative AI Toolkit added; page no longer says no guidance documents exist

· 1 page

What changed

  • The page said the State Bar of Georgia had published no informal guidance documents on AI. The State Bar's AI page links a Generative AI Toolkit (last updated July 9, 2026) from the Special Committee, which says it is guidance for members and is not intended to establish a formal advisory opinion. The page now says so; it still lists no ethics opinion or committee report. (1 page)

Source

State Bar of Georgia, Generative AI Toolkit PDF (gabar.org, read November 18, 2026)

Pages that carry this correction (1)

New York ethics, Harvey and Legora pages: NYSBA's October 31 guidelines, a reported Harvey stock-market listing discussion and Fieldfisher's Legora rollout are now covered

· 3 pages

What changed

  • The New York ethics page described only the April 6, 2024 Task Force report. It now also covers the guidelines the NYSBA House of Delegates approved on October 31, 2026 (billing, confidentiality and court filings), and says they are a committee document, not a formal ethics opinion. (1 page)
  • The Harvey valuation page said only that Harvey had announced no plans to go public. It still says that, and now adds that on November 6, 2026 Artificial Lawyer reported, citing its own sources, that Harvey has talked to advisors about a possible stock-market listing. Harvey has not confirmed it. (1 page)
  • The Legora page now covers Fieldfisher's rollout to about 1,400 people after a six-week pilot with 137 users, as reported by Irish Legal News. It is press coverage; no Fieldfisher or Legora release was read. (1 page)

Source

Legal Ethics Guidelines for Lawyers' Use of Artificial Intelligence, NYSBA Committee on Artificial Intelligence and Emerging Technologies, Version 7 (October 23, 2026), approved by the House of Delegates October 31, 2026 (read November 6, 2026); Artificial Lawyer, November 6, 2026, press report citing its own sources, not confirmed by Harvey (read November 6, 2026); Irish Legal News, November 6, 2026, press report on Fieldfisher and Legora, not a Fieldfisher or Legora release (read November 6, 2026)

Pages that carry this correction (3)

Pennsylvania pages: the Supreme Court's proposed AI notice for all court filers is now covered

· 2 pages

What changed

  • The page said that no timeline or proposal for a statewide attorney AI rule had been announced. That was out of date: on October 24, 2026 the Supreme Court's Advisory Committee on Artificial Intelligence published a proposed notice in the Pennsylvania Bulletin that would require every court filer to check the accuracy of generative AI output and certify that they did. It is a proposal only, with no rule number and no effective date, and comments are due by November 24, 2026. No statewide attorney AI rule is in force today. (2 pages)

Source

Pennsylvania Bulletin, vol. 56, no. 43, 56 Pa.B. 6888 (October 24, 2026): Proposed Notice Regarding the Use of Generative Artificial Intelligence by All Court Filers and Certification of Compliance (read November 6, 2026)

Pages that carry this correction (2)

Italian DPA v. OpenAI, A&O Shearman playbook and Stanford legal-research pages: claims without a primary source removed or qualified

· 3 pages

What changed

  • The Italian DPA v. OpenAI page called the EUR 15 million fine the first GDPR penalty against a generative AI company and, citing press reports, said the Rome court ruled on jurisdiction. The Garante's own page supports neither statement, so both were removed; the page now says we have not read the judgment and do not state its grounds. (1 page)
  • The A&O Shearman page described Linklaters' 20-person AI team and called the rollout 'production deployments handling real client work'. No source we read supports either statement (A&O Shearman's announcement lists four initial agent areas and a rollout plan, with no results), so the Linklaters passages were removed and the page now labels the staffing and implementation advice as our analysis. (1 page)
  • The Stanford study page gave Westlaw's accuracy as 42% only. The paper's abstract and summary say 42%, while its section 4 text says 41%; the page now shows 42% and notes the 41% figure. (1 page)

Source

Garante per la protezione dei dati personali, press release on ChatGPT (read November 6, 2026); A&O Shearman, announcement of the Harvey agents rollout, April 6, 2025 (read November 6, 2026); Magesh et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, arXiv 2405.20362 (read November 6, 2026)

Pages that carry this correction (3)

Wyoming page: federal General Order No. 25-01 added, unsourced claims removed

· 1 page

What changed

  • The page said the U.S. District Court for the District of Wyoming had issued no standing order on AI. In fact General Order No. 25-01, filed May 13, 2025, reminds litigants of Rule 11 and says citing hallucinated authority is strictly prohibited and subject to sanction. We added it and removed unsourced statements about the size of the Wyoming bar, rankings against other Tenth Circuit states, and predictions. (1 page)

Source

U.S. District Court for the District of Wyoming, General Order No. 25-01 on AI in the preparation of filings (read November 5, 2026); Wyoming State Bar homepage (read November 5, 2026); Wyoming Judicial Branch homepage (read November 5, 2026)

Pages that carry this correction (1)

Utah page: removed unsourced superlatives and 'no opinion' claims, added Garner v. Kadince

· 1 page

What changed

  • The page called Utah's sandbox the only one in the country and the most forward-thinking approach, and said the Utah Bar had no ethics opinion and no Utah judge had an AI order, without having read primary sources for those points. We removed those claims and now describe Standing Order No. 15 and the Bar's sandbox status as read. We added the Utah Court of Appeals opinion in Garner v. Kadince (2025 UT App 80), which sanctioned an attorney over fabricated AI citations and cited no Utah AI rule. (1 page)

Source

Utah Supreme Court Standing Order No. 15 (read November 5, 2026); Utah State Bar, Utah Legal Services Innovation Office page (read November 5, 2026); Garner v. Kadince, 2025 UT App 80, copy of the opinion (read November 5, 2026)

Pages that carry this correction (1)

Tennessee page: the AI task force is the Bar Association's, and predictions of coming rules removed

· 1 page

What changed

  • The page said Tennessee had convened a task force to study AI in the courts and that formal rules were likely within 12 to 18 months. The only task force page we could read is the Tennessee Bar Association's, which describes a bar group with no court role, deadline or recommendation, so we removed the court task force claim, the timeline and the prediction of coming rules. We also removed the count of judicial districts and rewrote the ethics section to rely on what the Whiting opinion says. The Tennessee court and Board of Professional Responsibility sites were unavailable to automated review, so the page now says we could not check them. (1 page)

Source

Tennessee Bar Association, AI Task Force (read November 5, 2026); U.S. Court of Appeals for the Sixth Circuit, Whiting v. City of Athens, opinion on sanctions, March 13, 2026 (read November 5, 2026); U.S. District Court, Middle District of Tennessee, administrative orders (read November 5, 2026)

Pages that carry this correction (1)

South Dakota page: unsourced bar-size, caseload and prediction claims removed

· 1 page

What changed

  • We re-read the South Dakota court, State Bar and federal court pages. None mentions AI, and the State Bar's ethics-opinions index (2019 to 2025) lists no AI opinion. We removed the attorney count, the claims about bar size and federal caseload, the number of judicial circuits, the statements about AI training data, and the prediction that South Dakota will adopt a rule. The page now describes only the pages we read and says we did not search judges' standing orders. (1 page)

Source

South Dakota Unified Judicial System, Supreme Court rules pages (read November 5, 2026); State Bar of South Dakota, ethics opinions index (read November 5, 2026); U.S. District Court, District of South Dakota, local rules and orders (read November 5, 2026)

Pages that carry this correction (1)

South Carolina court AI page corrected: the Supreme Court issued an interim policy in March 2025

· 1 page

What changed

  • The page said South Carolina had no supreme court order on AI. The Supreme Court of South Carolina in fact issued an Interim Policy on the Use of Generative AI on March 25, 2025, which applies to Judicial Branch personnel and reminds lawyers that they are responsible for the accuracy of their work product and for client confidentiality under Rule 407, SCACR. We also removed an unsourced attorney count, comparisons with other Fourth Circuit states, practice-area claims and predictions, and now state what we read on the Bar, federal court and Fourth Circuit sites. (1 page)

Source

Supreme Court of South Carolina, Interim Policy on the Use of Generative AI, Mar. 25, 2025 (read November 5, 2026); South Carolina Judicial Branch notice, Aug. 15, 2025 (read November 5, 2026); South Carolina Bar ethics advisory opinions index, page 1 (read November 5, 2026)

Pages that carry this correction (1)

Pennsylvania court AI page corrected against the Supreme Court's policy and the courts' own orders

· 1 page

What changed

  • The page now describes the Interim Policy as the Supreme Court approved it (September 9, 2025, effective December 8, 2025) and what the policy says about approved tools, disclosure to supervisors and human review. We removed a claimed prohibition on sensitive decisions, comparisons with other states, predictions about future rules, the claim that Western District judges have AI orders, and an incorrect statement that the Advisory Committee includes practitioners and technologists. We added the federal orders and the Butler County affidavit we read. (1 page)

Source

Supreme Court of Pennsylvania order approving the Interim Policy, Sept. 9, 2025 (read November 5, 2026); Pennsylvania Courts, AI Advisory Committee (read November 5, 2026); E.D. Pa. Standing Order Re: Artificial Intelligence, Judge Baylson, June 6, 2023 (read November 5, 2026); Butler County Affidavit Regarding Generative AI (read November 5, 2026)

Pages that carry this correction (1)

Oregon AI court disclosure page re-read against the Oregon State Bar, Court of Appeals and District of Oregon sources

· 1 page

What changed

  • We removed a reference to an Oregon State Bar Formal Opinion 2026-208 that we could not find on osbar.org, and unsourced statements about the Ninth Circuit and about individual District of Oregon judges. We corrected the Couvrette v. Wisnovsky figures to match the court's orders: a $15,500 sanction and an award of about $94,700 in attorney fees and costs ($110,204 in total), not an adverse costs order. We also replaced the claim that Oregon has no per-citation sanction valuation, since the Court of Appeals in Ringo used $500 per nonexistent case and $1,000 per fabricated quotation, and the District of Oregon adopted it in Couvrette. (1 page)

Source

Oregon State Bar, Formal Opinion No. 2025-205, Artificial Intelligence Tools (read November 5, 2026); Oregon Court of Appeals, Notice Regarding Court Imposition of Sanctions for Submission of Fabricated Authority Produced by AI (read November 5, 2026); D. Or., Couvrette v. Wisnovsky, Opinion and Order, ECF 215, December 12, 2025 (read November 5, 2026); D. Or., Couvrette v. Wisnovsky, Opinion and Order, ECF 225, March 23, 2026 (read November 5, 2026)

Pages that carry this correction (1)

Ohio AI court disclosure page re-read against the Ohio Supreme Court, Board of Professional Conduct and federal court sources

· 1 page

What changed

  • We removed the statements that a statewide Ohio AI disclosure rule was not imminent and that courts with educational resources usually adopt formal guidance within 12 to 24 months, because neither had a source. The page previously said the Board of Professional Conduct had issued no AI ethics guidance; it has published a nonbinding 2026 Ohio Ethics Guide, which we now describe. We also corrected the federal section: the Boyko and Greenberg orders we read ban AI in filings and Judge Graham's order requires a disclosure declaration, rather than all being disclosure requirements. (1 page)

Source

Supreme Court of Ohio, Artificial Intelligence Resource Library (read November 5, 2026); Ohio Board of Professional Conduct, Ohio Ethics Guide: Artificial Intelligence for Lawyers and Judicial Officers (read November 5, 2026); S.D. Ohio, Judge Graham, Standing Order on the Use of Generative Artificial Intelligence, October 21, 2025 (read November 5, 2026); N.D. Ohio, Judge Boyko, Standing Order on the Use of Generative AI (read November 5, 2026)

Pages that carry this correction (1)

North Dakota page: court system's AI notice detail added; unsourced bar-size and risk claims removed

· 1 page

What changed

  • The page said the North Dakota Supreme Court had not acted and described a small-bar risk theory. The Court System's June 1, 2026 notice states that disclosure is not required solely because AI was used but may be required under existing rules or orders in specific circumstances, and reserves sanction authority under N.D.R.Ct. 11.5. We removed the attorney headcount, the claim that a first sanction would be the state's first, and the claim about smaller training datasets, none of which had a primary source. (1 page)

Source

North Dakota Court System, Public Notice on the Use of AI, June 1, 2026 (read November 5, 2026); U.S. District Court, District of North Dakota, Local Rules effective October 6, 2025 (read November 5, 2026)

Pages that carry this correction (1)

New Mexico page: federal judge's AI disclosure order added; unsourced claims removed

· 1 page

What changed

  • The page said New Mexico had no AI disclosure requirement at any level and no individual judge orders. The U.S. District Court for the District of New Mexico has a standing order from Judge Margaret I. Strickland requiring disclosure of AI use and certification of accuracy in her cases, and the New Mexico Supreme Court has a permanent AI committee and has sanctioned an attorney over AI errors in a brief (the last two reported by secondary sources). We also removed an attorney headcount, comparisons with other Tenth Circuit states and an unsourced practice-area section. (1 page)

Source

D. N.M. Standing Order Establishing Disclosure and Certification Requirements for Generative AI, Judge Strickland (nmd.uscourts.gov) (read November 5, 2026); University of New Mexico Trustworthy AI, November 14, 2025 (read November 5, 2026); ABA Journal, August 24, 2026 (read November 5, 2026)

Pages that carry this correction (1)

Lawmatics free-trial guide: Essential does not include two-way SMS, Enterprise minimum and add-on pricing updated

· 1 page

What changed

  • The page said Essential includes SMS and that two-way SMS is 50 a month on Essential and Premium. The plan comparison read on November 5, 2026 shows no two-way SMS on Essential, 50 a month on Premium and 100 a month on Enterprise, with MMS as an add-on. The page also now gives the Enterprise user minimum (5), says Time and Billing is included on Enterprise, and names the lead-scoring add-on Merlin Qualify with outcome-based pricing instead of 'QualifyAI' with credit pricing. (1 page)

Source

Lawmatics pricing page, plan comparison (read November 5, 2026)

Pages that carry this correction (1)

Indiana: page said courts had adopted no AI-related rules; Supreme Court Order 26S-MS-8 took effect August 1, 2026

· 1 page

What changed

  • The page said Indiana had not adopted court rules reaching AI in legal practice. The Indiana Supreme Court's Order 26S-MS-8 (filed July 14, 2026, effective August 1, 2026) amended Trial Rule 11 so an attorney's signature certifies that any cited legal authority is authentic, and adopted Rules 11.1 (parties without an attorney) and 11.2 (abusive litigation). The page now describes the order. We found no Indiana rule requiring disclosure of AI use. (1 page)

Source

Indiana Supreme Court, Order Amending Rules of Trial Procedure, Cause No. 26S-MS-8 (in.gov/courts, read November 4, 2026)

Pages that carry this correction (1)

Georgia: State Bar AI committee named correctly and its reporting line added

· 1 page

What changed

  • The page called the State Bar of Georgia committee the Special Committee on Artificial Intelligence and Technology. The State Bar's own committee page names it the Artificial Intelligence and Emerging Technology committee and says it reports to the Executive Committee, the Board of Governors and the Supreme Court of Georgia. The page now uses that name and, as of November 4, 2026, still found no published opinion or report. (1 page)

Source

State Bar of Georgia, Artificial Intelligence and Emerging Technology committee page (gabar.org, read November 4, 2026)

Pages that carry this correction (1)

CoCounsel vs ChatGPT: ChatGPT prices without a vendor source removed, 'cites real cases' reworded

· 1 page

What changed

  • The page gave ChatGPT as $20 to $25 a month and ChatGPT Business as $20 per user per month billed annually. OpenAI's pricing and Business pages listed no dollar amounts when read on November 4, 2026, so the page no longer quotes a ChatGPT price. It also said CoCounsel 'cites real cases'; it now says CoCounsel's answers are grounded in the Westlaw database and link to sources that still have to be checked. (1 page)

Source

OpenAI, ChatGPT pricing page (chatgpt.com/pricing, read November 4, 2026): no plan amounts listed

Pages that carry this correction (1)

Vendor claims we had not checked: ChatGPT zero-retention, Claude Team and Enterprise contract terms, and SOC 2 as a given

· 3 pages

What changed

  • Three guides said ChatGPT Enterprise and Team offer zero-data-retention agreements, that Claude Team and Enterprise come with contractual guarantees, SOC 2 compliance and zero retention, that enterprise legal AI is SOC 2 Type II certified with zero retention by default, and that Claude Team is suitable for privileged work per Anthropic's published commitments. We have not reviewed OpenAI's or Anthropic's contracts for those plans. The pages now cite what we did read, that Anthropic's Commercial Terms say it may not train models on Customer Content, and tell firms to check retention, training and security-report terms in their own agreement. (3 pages)

Source

Anthropic Commercial Terms, section B: "Anthropic may not train models on Customer Content from Services" (read November 4, 2026)

Pages that carry this correction (3)

SOC 2 Type II stated as a fact for Harvey, Spellbook, Claude, ChatGPT, Relativity and Copilot, now attributed to what each vendor's page says

· 31 pages

What changed

  • Pages said Harvey is SOC 2 Type II certified or provides SOC 2 compliance with enterprise data agreements. Harvey's security page says it undergoes annual SOC 2 Type II and ISO 27001 audits and never trains on customer data by default (read November 4, 2026). We have not seen the audit report or any data agreement, so the pages now attribute each point to the security page. (7 pages)
  • Pages said Claude or Anthropic holds SOC 2 Type II certification, in places alongside 'does not sell data' or 'guarantees no training'. Anthropic's Privacy Center lists SOC 2 Type I and Type II among its certifications (read November 4, 2026); it says nothing about selling data. The pages now cite the Privacy Center and drop the sale-of-data wording. (11 pages)
  • Pages said Spellbook is SOC 2 Type II certified and does not train on client documents, that ChatGPT Team and Enterprise are SOC 2 Type II compliant, that Relativity aiR is FedRAMP authorized and SOC 2 certified, and that Microsoft Copilot carries SOC 2 Type II. Spellbook's page shows only a SOC 2 Type II Compliant badge; Relativity lists SOC 2 Type II for RelativityOne; Microsoft's Copilot privacy page does not mention SOC 2; OpenAI's page could not be read, so ChatGPT's status is marked unverified (all read November 4, 2026 or as dated on the page). (16 pages)

Source

Harvey, Security page (harvey.ai, read November 4, 2026): annual SOC 2 Type II and ISO 27001 audits; Spellbook, Security page (read November 4, 2026): shows a SOC 2 Type II Compliant badge; nothing on model training or encryption; Anthropic Privacy Center (read November 4, 2026): lists SOC 2 Type I and Type II among its certifications; Microsoft, Microsoft 365 Copilot privacy (learn.microsoft.com, read November 4, 2026): no SOC 2 mention; says Graph data is not used to train foundation LLMs

Pages that carry this correction (31)

SOC 2 Type II stated as a fact for Harvey and GC AI, now attributed to what each vendor's page says

· 4 pages

What changed

  • Pages said Harvey holds SOC 2 Type II certification, that its no-training commitment is contractual, and that it meets enterprise security requirements. Harvey's security page says it undergoes annual SOC 2 Type II and ISO 27001 audits and does not train on customer data (read November 4, 2026). We have not seen the audit report or Harvey's contract, so the pages now attribute each point to the security page. (3 pages)
  • The GC AI profile said the vendor lists SOC 2 Type II and SOC 3. GC AI's security page says it is SOC 2 Type II certified and shows SOC 3 only as a badge label, with zero-data-retention agreements "wherever feasible" (read November 4, 2026). The profile now says exactly that. (1 page)

Source

Harvey, Security page (harvey.ai, read November 4, 2026): annual SOC 2 Type II and ISO 27001 audits; does not use inputs, outputs or uploaded documents to train underlying models; GC AI, Security page (gc.ai, read November 4, 2026): "We are SOC 2 Type II certified"; SOC 3 shown only as a badge label; zero-data-retention agreements with LLM providers "wherever feasible"

Pages that carry this correction (4)

SOC 2 Type II stated as a fact for Spellbook and Claude, now attributed to what each vendor's page says

· 7 pages

What changed

  • Pages said Spellbook is SOC 2 Type II compliant, does not train on client documents and encrypts data, citing its privacy policy. Spellbook's security page shows a SOC 2 Type II Compliant badge and zero-data-retention agreements with OpenAI and Anthropic, but says nothing about model training or encryption, and we had not read a privacy-policy statement. The pages now report only what the security page shows (read November 4, 2026) and tell firms to get training and retention terms in writing. (6 pages)
  • The Harvey-versus-build-your-own-Claude page said the Claude API is SOC 2 Type II certified with enterprise data agreements. It now cites Anthropic's Privacy Center, which lists SOC 2 Type I & Type II for commercial products, and Anthropic's Commercial Terms on training (both read November 4, 2026). (1 page)
  • The Harvey versus Spellbook comparison gave both vendors SOC 2 Type II, no-training and encryption as settled facts. It now attributes each statement to the vendor's security page (read November 4, 2026) and says both are vendor statements, not contract terms. (1 page)

Source

Harvey, Security page (harvey.ai, read November 4, 2026): annual SOC 2 Type II and ISO 27001 audits; never trains on customer data by default; Spellbook, Security page (read November 4, 2026): shows SOC 2 Type II Compliant and HIPAA Compliant, zero data retention agreements with OpenAI and Anthropic; says nothing on model training or encryption; Anthropic Privacy Center, What certifications has Anthropic obtained? (read November 4, 2026): lists SOC 2 Type I & Type II for commercial products

Pages that carry this correction (7)

ChatGPT Enterprise SOC 2 Type II compliance listed as a plan inclusion without a vendor source, now marked unverified

· 1 page

What changed

  • The page listed SOC 2 Type II compliance as part of ChatGPT Enterprise without citing a vendor statement. No OpenAI statement on SOC 2 was available to cite, so the line now says the status is not verified and tells readers to ask OpenAI for its current SOC 2 report. (1 page)

Source

OpenAI, Enterprise privacy (openai.com), no SOC 2 statement available to cite on November 4, 2026

Pages that carry this correction (1)

Privilege and AI pages: removed a Kovel test no court has adopted, vendor claims we had not checked, an unnamed Chicago case and a judge count we could not source

· 4 pages

What changed

  • The guide said enterprise AI tools may qualify for Kovel protection if three conditions are met (necessity, a confidentiality agreement, reasonable precautions) and that this was a matter of confidential agreements alone. Kovel asks whether the helper's presence is necessary, or at least highly useful, for the lawyer's consultation with the client; the other two conditions are our own framework, and no court decision we read applies Kovel to an AI provider (Heppner mentions it only as a possibility). The page now says so and labels the three checks as this guide's own framework. (1 page)
  • The guide said AI vendors without a confidentiality agreement are strangers to the relationship and that courts make no exception for AI, and that courts have consistently held that sharing with third parties waives privilege, with no case cited. We removed the first claim; the second now says only what Heppner applied, the established rule that voluntarily sharing privileged information with a third party waives privilege. (1 page)
  • The guide said Harvey AI, CoCounsel and Lexis+ AI are built for this scenario, meet the contractual requirements and give the strongest privilege argument, citing no vendor document. We have not reviewed any vendor's contract, so the page now says to verify the terms in your own agreement and no longer names a safest or strongest tool. (1 page)
  • The guide's decision tree said enterprise tools with data protection agreements mean privilege may be preserved, said client consent removes any risk that a client claims they did not authorize third-party access, and closed on enterprise AI with Kovel doctrine protections as the safer path. No case we read decides the enterprise question, so the page now says that, says telling clients up front answers a later claim of non-authorization, and calls a written confidentiality agreement the more defensible path while noting no court has decided it. (1 page)
  • The page said that under the Kovel doctrine and its modern extensions, enterprise AI vendors can qualify as agents of the attorney. We found no decision that says so. It now says whether a vendor under contract can count as the lawyer's agent is untested for AI, and that Heppner mentions it only as an open possibility. (1 page)
  • The page opened with a March 2026 Chicago firm, an AI agent, a waiver motion and a court that agreed to hear it, with no case name or docket, and said this was not hypothetical. We could not source it, so it is now an explicit scenario and says we know of no court decision on that argument. (1 page)
  • The page said over 30 federal judges had adopted standing orders on AI work product and disclosure, and that judicial consensus had crystallized around three requirements. We have no list or count behind the number, and the cases on the page differ, so the number is gone and the page now describes disclosure, verification and accountability as recurring themes on which courts differ. (1 page)

Source

United States v. Heppner, No. 25 Cr. 503 (S.D.N.Y.), ECF No. 27, order of February 17, 2026 (read November 4, 2026); Morgan v. V2X, Inc., No. 1:25-cv-01991-SKC-MDB (D. Colo.), ECF No. 65, order of March 30, 2026 (read November 4, 2026); United States v. Kovel, 296 F.2d 918 (2d Cir. 1961) (read November 4, 2026)

Pages that carry this correction (4)

Attorney-client privilege and AI guide: removed a case claim we could not source and replaced an empty list of court decisions with three decisions read from the courts' own orders

· 1 page

What changed

  • The guide named a "Grand Jury AI subpoena" scenario as if it were a case and said a government subpoena to an AI vendor leaves no basis to quash. We found no court decision by that name or on that scenario, so the sentence and the matching FAQ answer now say the question is unresolved and state no rule. (1 page)
  • The section "What Courts Are Saying in 2025-2026" promised several court decisions but listed none, and described an emerging standard and a consistent trend without naming a source. It now lists United States v. Heppner, Warner v. Gilbarco and Morgan v. V2X, each read from the court's own order, with the judge's office (Heppner is a district judge's ruling; Warner and Morgan are magistrate judges' orders) and what each decided. (1 page)

Source

Warner v. Gilbarco, Inc., No. 2:24-cv-12333 (E.D. Mich.), ECF No. 94, order of February 10, 2026 (read November 4, 2026); Morgan v. V2X, Inc., No. 1:25-cv-01991-SKC-MDB (D. Colo.), ECF No. 65, order of March 30, 2026 (read November 4, 2026)

Pages that carry this correction (1)

Unsourced 'partner time' hour estimates removed on 2 guides

· 2 pages

What changed

  • These pages gave hour estimates of partner time (for example 80-120 hours in year one of an AI rollout, 20-40 hours to build a template library, 30-40 hours of governance setup). No source supported any of them, so the pages now say we have not measured the partner time and tell the reader to time a pilot. (2 pages)

Source

Thomson Reuters Institute, Making the most of AI's potential time savings (September 12, 2024), the dated primary on lawyer time savings; none of the removed partner-time figures has a primary, so none was replaced by a number

Pages that carry this correction (2)

Heppner and privilege pages: corrected how the case began, who the defendant was, the case name and year, and removed enterprise rules, a judge count and vendor claims we could not source

· 29 pages

What changed

  • The Heppner explainer said prosecutors got the defendant's Claude chat logs by subpoenaing Anthropic and that Anthropic complied. The opinion (ECF No. 27) says the FBI seized about 31 documents from the defendant's home under a search warrant and does not address a subpoena to the provider. The page now says so and keeps only what the opinion notes: Anthropic's privacy policy allows disclosure to governmental regulatory authorities. (1 page)
  • The Heppner explainer put a quotation about an AI consultation structured by counsel in the judge's mouth and said the opinion spelled out requirements in a Kovel analysis. The opinion has no such quote and no such analysis: it says only that Claude might arguably have acted as the lawyer's agent if counsel had directed its use. The page now says that and labels the requirements as this guide's own framework. (1 page)
  • Three guides described Heppner as a lawyer who put client documents into ChatGPT, with ethics complaints and client notices. The defendant was a criminal defendant, not a lawyer, he used Claude on his own, and the opinion says nothing about ethics complaints or notices. The guides now describe the case as the opinion does. (3 pages)
  • Five guides called the case by a name that does not exist and one gave the wrong year, and one said multiple commentators had applied the case. The case is United States v. Heppner, No. 25 Cr. 503 (S.D.N.Y.), decided February 2026. We corrected the name and year and removed the unsourced reference to commentators; two sentences that credited the ruling with a distinction between consumer and enterprise tools were rewritten as well. (5 pages)
  • Several pages stretched one ruling, a defendant's own use of consumer Claude, into a rule for every client conversation, every public AI platform, every free chatbot, lawyers' research history and tax memos. They now say what the opinion held and that it did not decide enterprise tools or a firm's own use. (6 pages)
  • Guides on Claude, Copilot, Bedrock, Vertex and GitHub Copilot said enterprise tiers satisfy Heppner, sit on the right side of a line it draws, or meet an enterprise threshold, and one said an enterprise tool is functionally equivalent to a Kovel expert. The opinion does not decide enterprise tools and draws no such line. The pages now say that enterprise terms give a stronger confidentiality argument, that no court has ruled, and that the Kovel comparison and the split in the legal community are our own reading. (14 pages)
  • Three Project Deal guides said Heppner makes agent-to-agent transaction records unprivileged and spoke of a Heppner gap. Heppner concerned one defendant and Claude; it did not address agent-to-agent transactions, so the pages now call that question untested. (3 pages)
  • Three court-filing guides said over 30 federal judges require AI disclosure. We had no count and no list behind that figure, so the pages now say some federal judges do and that we have not counted them. (3 pages)
  • The confidentiality-waivers guide said ChatGPT Enterprise, Claude, Harvey and Relativity all offer contractual guarantees against retention and training, and that each must hold SOC 2 Type II certification. We have not reviewed any vendor's contract, so the page now says to check the terms in your own agreement and treats an independent security report as a baseline to ask for. (1 page)

Source

United States v. Heppner, No. 25 Cr. 503 (S.D.N.Y.), ECF No. 27, order of February 17, 2026 (read November 4, 2026)

Pages that carry this correction (29)

Heppner privilege ruling dated February 17, 2026 (not 2024) on the attorney-client privilege and AI guide

· 1 page

What changed

  • The guide dated the Heppner privilege ruling to 2024 and called it In re Heppner. The opinion is United States v. Heppner (S.D.N.Y.), a memorandum filed February 17, 2026; the dates and the case name now match the opinion. (1 page)

Source

United States v. Heppner, 25 Cr. 503 (JSR), ECF No. 27, memorandum of February 17, 2026 (read November 4, 2026)

Pages that carry this correction (1)

Attorney-client privilege and AI guide: the Heppner case described as the court's opinion describes it (a defendant's own use of Claude, with confidentiality resting on the provider's privacy policy)

· 1 page

What changed

  • The guide said the ruling found that consumer AI tools are third-party disclosure because inputs may be used for training, and described a lawyer pasting client communications into ChatGPT. The court held that a defendant's own exchanges with Claude were not confidential, relying on the provider's written privacy policy (it collects inputs and outputs, uses them to train the model and may disclose them to third parties, including governmental authorities); it did not address ChatGPT or a lawyer's use. (1 page)
  • The guide described the case as one about privileged information entered into a consumer AI tool. The defendant was not a lawyer: he used Claude on his own, without his counsel's direction, and the court also held the exchanges were not work product. (1 page)
  • The guide called the theory not novel, headed the ruling section as if consumer AI waives privilege as a rule, and closed that consumer tools waive privilege because they are unsecured third parties. The court called the question one of first impression nationwide; the section title and bottom line now state what was decided and what was not. (1 page)

Source

United States v. Heppner, 25 Cr. 503 (JSR), ECF No. 27, memorandum of February 17, 2026, read in full November 4, 2026

Pages that carry this correction (1)

Two claims with no primary source removed: an estimated Freshfields x Anthropic contract value and the languages Luminance Eve is said to process

· 2 pages

What changed

  • The guide on replicating the Freshfields x Anthropic deal said the contract value was likely in the $15-30M+/year range. The announcement discloses no contract value and we have no source for that estimate, so the page now says the value was not disclosed. (1 page)
  • The Eve by Luminance review said Eve processes contracts in German, Japanese, Portuguese and English consistently. Luminance publishes no list of supported languages that we could confirm, so the page now tells readers to ask Luminance for the list and test their own contracts. (1 page)

Source

Freshfields, announcement of the Anthropic collaboration (April 23, 2026), which gives scope and lawyer and office counts and no contract value; Luminance, Luna Crescent page, read November 4, 2026; it describes the model and publishes no list or count of supported languages

Pages that carry this correction (2)

Billing, discovery and vendor lines we could not source: Florida 24-1 on charging for AI, Aérospatiale, ABA 512 fee wording, vendor contract terms and two absolute statements

· 9 pages

What changed

  • Several billing pages said Florida Bar Opinion 24-1 bans charging a client for AI tools as a separate line item and that AI is overhead. The opinion says a lawyer may charge the client the actual cost of generative AI if the client is told in advance, preferably in writing, and the charge is reasonable and not duplicative of overhead. Only periodic charges that cannot be tied to a specific client are overhead and may not be prorated. It also says efficiency must not lead to falsely inflated claims of time. The pages now say that. (3 pages)
  • Two guides described a settled view among bars, attributed to ABA Formal Opinion 512, on billing for time AI saves. We did not re-check that opinion's fee section for this change. The pages now cite only what we did read: Florida 24-1 says efficiency must not result in falsely inflated claims of time and suggests contingent or flat fees for specific services. (2 pages)
  • The cross-border e-discovery page said US courts have uniformly treated the Hague Evidence Convention as optional and that EU courts take the opposite view. We read the Aérospatiale syllabus: the Convention does not provide exclusive or mandatory procedures, and courts decide case by case whether to try it first. We have no source on EU courts, so that sentence is gone. (1 page)
  • A Claude Mythos article said consumer ChatGPT, Claude's free tier and most low-cost paid plans lack the contract terms the Morgan v. V2X order requires, and a vendor checklist called SOC 2 Type II table stakes for Harvey AI, CoCounsel and Luminance. We have not read those vendors' plan contracts or SOC 2 reports. The pages now say so and tell readers to check each plan's terms or report. (2 pages)
  • Three court-disclosure pages said no lawyer has been sanctioned for disclosing AI use, and one said state courts are following the federal courts, naming three states. We have no survey of sanctions or of state court rules behind either, so both statements were replaced with advice to check each court's own rules. (3 pages)

Source

Florida Bar Ethics Opinion 24-1, billing and confidentiality sections (floridabar.org/etopinions/opinion-24-1/, read November 4, 2026); Société Nationale Industrielle Aérospatiale v. U.S. District Court, 482 U.S. 522 (1987), syllabus (Cornell LII, read November 4, 2026)

Pages that carry this correction (9)

Maryland and Delaware AI ethics pages: task force name and Delaware source corrected against the primaries

· 2 pages

What changed

  • The Maryland page named the author of the April 2025 overview the MSBA's AI and Legal Technology Task Force. The MSBA's own page names it the Artificial Intelligence Task Force; the page now uses that name. (1 page)
  • The Delaware page dated the Supreme Court's interim AI policy October 21 and 22, 2024 and linked a PDF as the policy text. That PDF is the court's press release of October 22, 2024, which says the court adopted the policy that day and says the order and policy are on the court's website. The page now dates it October 22, 2024 and calls the source a press release. (1 page)

Source

MSBA, An Overview of Ethical Considerations for Attorney Use of Generative Artificial Intelligence Technologies (store page, author named as the MSBA Artificial Intelligence Task Force), read November 4, 2026; Delaware Supreme Court press release of October 22, 2024 (the document the page linked as the policy text), read November 4, 2026

Pages that carry this correction (2)

Colorado AI law: eight pages still named the repealed SB 24-205 as current or in effect, and one quoted a $20,000 penalty with no primary

· 8 pages

What changed

  • The Utah page named the AI Act, SB 24-205, as Colorado's current regulation. That act was repealed and reenacted by SB 26-189, which applies from January 1, 2027, so the page now names SB 26-189. (1 page)
  • The Wyoming page said Colorado has statutory AI obligations through the AI Act (SB 24-205), which contradicted the rest of the page. It now names SB 26-189, applying from January 1, 2027. (1 page)
  • The New Mexico page described Colorado's AI Act (SB 24-205) as creating statutory obligations. It now names SB 26-189, which replaced it and applies from January 1, 2027. (1 page)
  • The April 2026 page said the Colorado AI Act is already in effect. It is not: the replacement law, SB 26-189, applies to consequential decisions made on or after January 1, 2027, and the page now says so. (1 page)
  • The board reporting guide listed the Colorado AI Act among active regulations. It now cites SB 26-189, which applies from January 1, 2027. (1 page)
  • The AI regulation hub and the Oregon page compared against Colorado's SB 205. They now name SB 26-189, which replaced SB 24-205. (2 pages)
  • The Colorado compliance guide quoted a penalty of $20,000 per violation that the Attorney General can seek. SB 26-189 states no amount and we have no primary source for that figure, so the page no longer quotes one. (1 page)

Source

Colorado General Assembly, SB 26-189 Final Act (Sec. 1 repeals and reenacts part 17 of article 1 of title 6; Sec. 5: applies to consequential decisions made on or after January 1, 2027), read November 4, 2026

Pages that carry this correction (8)

Legal AI Vendor Directory: seven price ranges and an '80% of the noise' figure with no source replaced by vendor pages or removed

· 1 page

What changed

  • The directory gave per-user and per-year price ranges for Claude Enterprise, Spellbook, Luminance, Relativity, Lex Machina, Smith.ai and OneTrust with no source, and said three filters eliminate 80% of the noise. Claude Enterprise now carries the seat price from Anthropic's pricing page and Smith.ai the plans on its own page; Spellbook, Lex Machina, Relativity and OneTrust show no price on their pages, Luminance's pricing page returned 404, so those entries now say to ask for a quote. The 80% figure was removed because we never measured it. (1 page)

Source

Anthropic, Claude pricing page, Enterprise self-serve at US$20 per seat per month billed annually plus usage at API rates, 20-seat minimum (read November 3, 2026); Smith.ai pricing page, human-answered plans at $300, $810 and $2,100 per month (read November 3, 2026); Spellbook pricing page, custom pricing with a demo request (read November 3, 2026); Lex Machina pricing page, a demo request and no price (read November 3, 2026)

Pages that carry this correction (1)

Vendor data-handling claims (SOC 2, training on client data, on-premise, FedRAMP, HIPAA) attributed to each vendor's own page or removed on 7 more comparison pages

· 7 pages

What changed

  • Harvey: replaced 'holds SOC 2 Type II certification, does not train on client data and offers enterprise data agreements' with what its security page says (never trains on customer data by default, annual SOC 2 Type II and ISO 27001 audits). The page states no default retention period, so the pages tell readers to ask for it in writing. (4 pages)
  • Lexis+ AI and Westlaw AI: removed 'operates under enterprise data agreements, no training on queries, isolated from consumer tools'. LexisNexis says it never uses customer data to train AI models and keeps retention policies; Thomson Reuters' no-training and zero-retention statements are about CoCounsel only, and we found none for Westlaw AI. (5 pages)
  • Relativity aiR: removed 'FedRAMP authorized, SOC 2 Type II, HIPAA compliant' as a blanket claim and the ranking above CoCounsel. Relativity lists SOC 2 Type II, SOC 3, ISO/IEC 27001 and 27018 and HIPAA compliance for RelativityOne and says its FedRAMP Moderate authorization applies to Relativity aiR Government. (2 pages)
  • Spellbook and Luminance: removed 'SOC 2 Type II certified and does not train on client documents' for Spellbook and 'on-premise deployment' for Luminance. Spellbook's page shows a SOC 2 Type II badge, zero-data-retention agreements with OpenAI and Anthropic and Canada/US data centers and does not say whether it trains; Luminance's page lists regular SOC 2 Type II examinations and no on-premise option. (1 page)
  • Claude: removed 'SOC 2 Type II certified' and 'does not sell data' and the claim that Claude's data handling is more transparent than Lexis+ AI's. Anthropic's pricing page lists model training as none by default on Team and Enterprise and states no SOC 2 attestation. (1 page)

Source

Harvey, security page (read November 3, 2026); LexisNexis, Lexis+ with Protégé page (read November 3, 2026); Relativity, aiR page (read November 3, 2026); Thomson Reuters, CoCounsel page (read November 3, 2026); Spellbook, security page (read November 3, 2026); Luminance, security page (read November 3, 2026); Anthropic, pricing page (read November 3, 2026)

Pages that carry this correction (7)

Vendor data-handling claims (training on client data, SOC 2, residency, on-premise, retention) attributed to each vendor's own page or removed on 11 pages

· 11 pages

What changed

  • Removed the claims that Legora is 'EU-native' or satisfies GDPR Article 28 by design and that Harvey's US infrastructure creates compliance exposure. Legora's page names no hosting region and says customers control where data is stored; Harvey's page says it offers processing in the EU and Switzerland or Australia. The pages now say that and tell readers to ask each vendor for its region and DPA. (2 pages)
  • Harvey: replaced 'certified' and 'does not train' with what its security page says (never trains on customer data by default, bespoke model only on request, annual SOC 2 Type II and ISO 27001 audits). Removed data-isolation, access-control, audit-logging, incident-response and 'stays within the firm's environment' statements that page does not make. (3 pages)
  • Removed the claim that Luminance offers on-premise deployment and 'edges ahead' on data residency; its security page does not state either. The page now reports what Luminance does say (SOC 2 Type II examinations, AES-256 and TLS 1.2 or higher, AWS hosting, customer learning kept private to each organization). (1 page)
  • Spellbook: removed 'doesn't train on your data' and 'trained on common business contract types'. Its security page shows zero-data-retention agreements with OpenAI and Anthropic, a SOC 2 Type II badge and Canada/US data centers, and does not say whether customer data trains models. (1 page)
  • CoCounsel and Westlaw AI: the no-training and zero-retention statements on Thomson Reuters' page are about CoCounsel only; the page no longer says Westlaw AI handles data identically. (1 page)
  • Claude and Copilot: removed 'SOC 2 Type II certified' for Claude (no statement found on Anthropic's pages), 'does not sell data', and 'data stays within your Microsoft 365 tenant' for Copilot (Microsoft says LLM calls can be routed to other regions at high utilization and Anthropic models inside Copilot are outside the EU Data Boundary). Both now cite the vendor page and date. (1 page)
  • Removed 'data isn't used for training on paid plans' (Anthropic lists training as opt-out on Pro and Max), 'paid enterprise tiers with zero-retention policies are required' and 'enterprise pricing reflects firm-specific model training'; none has a vendor source. (4 pages)

Source

Harvey, security page (read November 3, 2026); Legora, security page (read November 3, 2026); Luminance, security page (read November 3, 2026); Spellbook, security page (read November 3, 2026); Thomson Reuters, CoCounsel page (read November 3, 2026); Anthropic, pricing page (read November 3, 2026); Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot (read November 3, 2026)

Pages that carry this correction (11)

Vendor and model quality superlatives ('not close', 'most capable', 'unmatched', 'outperforms') reworded where no benchmark supports them

· 17 pages

What changed

  • These pages said Claude's legal writing is not close to other tools, that it produces the most polished legal prose, that it is noticeably better than Harvey in side-by-side tests, that it is the most capable AI for legal work, that ChatGPT does more than any other tool and has an unmatched custom GPT ecosystem, and that Claude teaches better than any textbook. No published test measures these; the Vals report did not include Claude or ChatGPT. The pages now describe what the tools do and cost and say that no independent comparison was found. (5 pages)
  • These pages said Harvey outperforms every other platform or is significantly better than Copilot, that Luminance beats Harvey, that Claude outperforms Copilot or beats Westlaw AI, that Kira's extraction is unmatched, that CoCounsel beats ChatGPT 'and it's not close', that Copilot beats most alternatives on copyright indemnification, and that a narrowly scoped agent 'will outperform' a generic one every time. The Vals report, the only dated head-to-head we found, covers four tools and none of the other products named; the Stanford study tested other research tools. The pages now state what each source measured and that no independent benchmark compares the products. (9 pages)
  • These pages called products 'the best value in legal AI' or among the best values without a comparison of value. They now state the published price or that pricing is quote-based and that we found no independent value benchmark. (5 pages)

Source

Vals Legal AI Report (updated February 27, 2025, read November 3, 2026) tested CoCounsel, Vincent AI, Harvey Assistant and Oliver, and did not test Claude, ChatGPT, Legora, Spellbook, Luminance, Kira or Microsoft Copilot; Magesh et al., Stanford RegLab and HAI (arXiv 2405.20362, read November 3, 2026) found Lexis+ AI, Westlaw AI-Assisted Research and Ask Practical Law AI hallucinated 17% to 33% of the time, and did not test CoCounsel, Harvey, Claude or ChatGPT

Pages that carry this correction (17)

Vendor and model quality superlatives ('dominates', 'wins', 'superior', 'better choice', 'only option') reworded where no benchmark supports them, second batch

· 18 pages

What changed

  • These pages named a winner or called a product the superior, stronger or better choice: Harvey over Westlaw AI and over Lexis+ Protege, Relativity aiR on security, CoCounsel over Westlaw AI for multi-task work, and Harvey and Legora by market. The Vals reports and the Stanford study did not test most of these pairs, and Relativity's trust page publishes no ranking. The pages now say what each product is built for and that no independent benchmark ranks them. (6 pages)
  • These pages said products or firms dominate or lead a category: Everlaw in e-discovery, Relativity and Nuix in review and processing, firms by deployment scale, innovation and AI governance, DISCO against Relativity, and Relativity's app marketplace. We found no independent market-share or benchmark data for them, so the pages describe what each is built for or say no ranking exists. (4 pages)
  • These pages said Claude wins, dominates or is superior to Copilot, Lexis+ AI, CoCounsel and Perplexity, that it is the best AI for criminal defense, that Anthropic's Enterprise retention terms are the only option that settles privilege, and that one pricing scenario was the clear winner. The Vals legal research study tested ChatGPT but not Claude, so no independent test supports these rankings. The pages now give our editorial pick as such, or the price and terms as published. (8 pages)

Source

Vals Legal AI Report (updated February 27, 2025, read November 3, 2026) tested CoCounsel, Vincent AI, Harvey Assistant and Oliver; Lexis+ AI withdrew; Vals legal research study (updated October 14, 2025, read November 3, 2026) tested Alexi, Counsel Stack, Midpage and ChatGPT, and did not test Claude, Harvey, CoCounsel, Westlaw AI or Lexis+ AI; Magesh et al., Stanford RegLab and HAI (arXiv 2405.20362, read November 3, 2026) found Lexis+ AI, Westlaw AI-Assisted Research and Ask Practical Law AI hallucinated 17% to 33% of the time, and did not test CoCounsel, Harvey, Claude or ChatGPT; Relativity trust page (read November 3, 2026) offers a FedRAMP-certified platform, Relativity aiR Government; it publishes no ranking against other vendors

Pages that carry this correction (18)

AI statute-of-limitations tracking: malpractice cost, premium and LawToolBox price figures with no source removed

· 1 page

What changed

  • The malpractice section said the average malpractice claim costs $30,000-$50,000 in defense costs, calendar-related claims $50,000-$150,000, that a missed deadline raises premiums 25-50%, and that LawToolBox costs $50-100 per user per month. No source supported any of these figures, and LawToolBox's own pricing page lists lower per-user monthly prices that fall with the number of users, so the page now gives no figures and tells the reader to ask their insurer and the vendor for written quotes. (1 page)

Source

LawToolBox pricing page, per-user monthly prices by number of users (read November 3, 2026)

Pages that carry this correction (1)

Washington state-court page: 'a task force is active and multiple AI bills are progressing' withdrawn

· 1 page

What changed

  • The page said a task force is active, multiple AI bills are progressing and formal court rules are expected, with no source. What we read: the Board for Judicial Administration heard an AI presentation on June 21, 2024, and Senate Bill 6073 (2024) would have required AI disclosure and a citation-verification certification in filings (its outcome was not confirmed). The page now says we could not confirm a current task force or pending bill, and lists the eight RPCs Advisory Opinion 2025-05 applies, including fees (RPC 1.5). (1 page)

Source

Washington State Bar Association, Advisory Opinion 2025-05, Artificial Intelligence-Enabled Tools in Law Practice (read November 3, 2026); Washington Senate Bill 6073, 2024 regular session (leg.wa.gov) (read November 3, 2026); Washington Board for Judicial Administration, meeting minutes, June 21, 2024 (courts.wa.gov) (read November 3, 2026)

Pages that carry this correction (1)

Virginia state-court page: HB1642 description and 'oversight mechanisms, reporting requirements' claims withdrawn

· 1 page

What changed

  • The page said HB1642 covers judicial decision-making, case management and court administration and that the court AI program has oversight mechanisms and reporting requirements. No source we found supported those details, and we rely on secondary reporting for the enacted text. The page now reports HB1642 as the National Court Reporters Association does (human oversight of AI in judicial processes, effective July 1, 2025) and describes the court system's authorized-use page and the Supreme Court of Virginia's Rules for Generative AI Use, which govern court personnel and are not an attorney filing rule. (1 page)

Source

Virginia court system, Authorized Use of Artificial Intelligence in Virginia's Judicial System (vacourts.gov) (read November 3, 2026); Supreme Court of Virginia, Rules for Generative AI Use, effective June 6, 2024, last amended May 7, 2025 (read November 3, 2026); Supreme Court of Virginia, Legal Ethics Opinion 1901, approved November 24, 2025 (read November 3, 2026); National Court Reporters Association, report on HB1642 (thejcr.com), April 7, 2025 (read November 3, 2026)

Pages that carry this correction (1)

Texas state-court page: TRAIGA does not impose AI documentation or accountability duties on law firms

· 1 page

What changed

  • The page said TRAIGA creates transparency, documentation and accountability duties that apply to law firms. In the enrolled text, effective January 1, 2026, the disclosure duty (Sec. 552.051) applies to governmental agencies and health care providers, the prohibitions cover specific harmful uses, the attorney general enforces, and nothing requires a law firm to document its AI systems or disclose AI use in filings. The page now says that. (1 page)
  • Removed claims with no source: that Judge Starr's order 'became a national model', that Texas has over 3,000 state court judges, that the State Bar's AI Task Force is active and the court system has published guidance, and that candor and competence duties sit in the Rules of Civil Procedure (they sit in the Disciplinary Rules). (1 page)

Source

Texas H.B. 149, Texas Responsible Artificial Intelligence Governance Act, enrolled text (capitol.texas.gov) (read November 3, 2026); U.S. District Court N.D. Tex., Local Civil Rules effective September 1, 2026, LR 7.2(f) (read November 3, 2026)

Pages that carry this correction (1)

Oklahoma state-court page: 'the Supreme Court hasn't acted' and judge-by-judge disclosure claims withdrawn

· 1 page

What changed

  • Removed the statements that the Oklahoma Supreme Court had not acted and that individual state judges had begun issuing disclosure requirements. The Southwest Ledger quotes the Supreme Court's March 23, 2026 decision: 'Disclosure of AI assistance shall not be required in a legal filing' (we have not read the order itself). We found no source for any state district judge's AI order, so the page now says so. (1 page)
  • Removed 'No Oklahoma judge has penalized an attorney for voluntarily disclosing AI use', which had no source. (1 page)
  • Removed the prediction that Oklahoma statewide rules will come. (1 page)

Source

Oklahoma Court of Criminal Appeals, Order 2026 OK CR 7, adopting Rule 1.17 (okcca.net), February 18, 2026 (read November 3, 2026); Oklahoma Court of Criminal Appeals, Rule 1.17 Use of Generative Artificial Intelligence (okcca.net) (read November 3, 2026); Southwest Ledger, 'State Supreme Court advises caution in use of AI' (secondary report quoting the Supreme Court) (read November 3, 2026)

Pages that carry this correction (1)

New Hampshire state-court page: 'no Bar guidance' withdrawn; NHBA articles and D.N.H. local rules described

· 1 page

What changed

  • The page said the New Hampshire Bar had issued no AI guidance and the Supreme Court had not acted. The Bar's Ethics Committee published two advisory AI articles (May and September 2024); the District of New Hampshire's December 2025 local rules contain no AI provision; the Supreme Court's December 2025 task force on judicial AI use is reported only in search results (courts.nh.gov was not available to our checks). The attorney count and the Massachusetts and First Circuit comparisons were removed. (1 page)

Source

New Hampshire Bar Association Ethics Committee, Using Artificial Intelligence in Practice, May 15, 2024 (nhbar.org) (read November 3, 2026); New Hampshire Bar Association Ethics Committee, Drafting Documents with Artificial Intelligence, September 10, 2024 (nhbar.org) (read November 3, 2026); U.S. District Court, District of New Hampshire, Local Rules effective December 1, 2025 (nhd.uscourts.gov) (read November 3, 2026)

Pages that carry this correction (1)

Nevada state-court page: unsourced 'no rule at any level' and 'no standing orders' absolutes replaced with what was read

· 1 page

What changed

  • The page asserted that Nevada has no AI rule at any level, that the Supreme Court and State Bar had issued nothing and that no judge has an AI order. None of that was verified against rule text. The page now says 'we found no' and states the limit of what we found. (1 page)

Source

State Bar of Nevada, ethics opinions index (nvbar.org; no readable opinion text loaded) (read November 3, 2026); U.S. District Court, District of Nevada, local rules page (no readable rule text loaded) (read November 3, 2026)

Pages that carry this correction (1)

Nebraska state-court page: 'no federal local rule' and 'silence is complete' withdrawn; NECivR 7.1(d)(2) and the March 2026 referral added

· 1 page

What changed

  • The page said the District of Nebraska had no AI local rule and that Nebraska's silence was complete. NECivR 7.1(d)(2) (effective December 1, 2025) requires parties using generative AI to verify their filings, and the Nebraska Supreme Court referred an attorney to the Counsel for Discipline in March 2026 over a brief with fictitious cases (news report; opinion not read). The attorney count, the small-bar comparison and the Eighth Circuit sanctions claim were removed. (1 page)

Source

U.S. District Court, District of Nebraska, Local Civil Rules effective December 1, 2025, NECivR 7.1(d)(2) (ned.uscourts.gov) (read November 3, 2026); Nebraska Public Media, report on the Nebraska Supreme Court's March 20, 2026 opinion on AI-authored filings (nebraskapublicmedia.org) (read November 3, 2026); Vermont Judiciary Committee on Artificial Intelligence and the Courts, First Annual Report, March 1, 2025 (vtcourts.gov) (read November 3, 2026)

Pages that carry this correction (1)

Montana state-court page: 'complete inaction', 'no regulatory activity' and bar size withdrawn; Rule 3(G) and proposed Rule 901(c) described

· 1 page

What changed

  • The page said Montana had taken no action, while the Fourth Judicial District requires disclosure under Rule 3(G) (now described from the rules PDF) and the Montana Supreme Court has proposed a Rule of Evidence 901(c) on AI-altered evidence (State Bar notices; final status not confirmed). The attorney count was replaced by the State Bar's reported 4,245 active attorneys, and the Ninth Circuit comparison was removed. (1 page)

Source

Montana Fourth Judicial District, Rules of Practice, rev. October 31, 2025, Rule 3(G) (missoulacounty.gov) (read November 3, 2026); State Bar of Montana, notice on the Supreme Court's proposed Rule of Evidence 901(c) (montanabar.org) (read November 3, 2026)

Pages that carry this correction (1)

Missouri state-court page: 'only the 20th Circuit has an AI rule' and 'nothing is pending' claims withdrawn

· 1 page

What changed

  • The page said only the 20th Judicial Circuit has an AI disclosure rule and that nothing is pending. HeplerBroom (March 24, 2026) cites local rules on generative AI disclosure in the 7th, 17th, 20th and 46th Circuits. The rule texts are not quoted on the page, so it now attributes the account to HeplerBroom and does not claim an exclusive 20th Circuit rule, that nothing is pending, or that the rule will spread. (1 page)

Source

HeplerBroom, 'Tips for Avoiding the Pitfalls of AI in Legal Practice: Missouri's Evolving Rules on AI Disclosure' (secondary), March 24, 2026 (read November 3, 2026)

Pages that carry this correction (1)

Mississippi state-court page: 'judge was transparent', 'court staff used AI' and 'no Bar opinion' withdrawn

· 1 page

What changed

  • The page said a judge disclosed that court staff used AI and called the disclosure transparent. The documents we could read show a July 20, 2025 order that misnamed parties and misquoted state law, replaced with a corrected version, and an October 21, 2025 reply calling the original a draft that should not have been docketed; we have not seen the letter (grassley.senate.gov was not available to our checks), so the page no longer says who used which tool. The Mississippi Bar's AI guide points to an ethics opinion, so the 'no opinion' line was removed. Unsourced counts and the Texas statute claim were removed. (1 page)

Source

U.S. Senate Committee on the Judiciary, press release on Judge Wingate's July 20, 2025 order, October 6, 2025 (judiciary.senate.gov) (read November 3, 2026); WLBT, report on Judge Wingate's October 21, 2025 response (wlbt.com) (read November 3, 2026); U.S. District Court, N.D. Texas, Judge Brantley Starr page, Local Civil Rule 7.2(f) (txnd.uscourts.gov) (read November 3, 2026)

Pages that carry this correction (1)

Minnesota state-court page: 'pending legislation on legal practice' and 'active task force' withdrawn

· 1 page

What changed

  • The page said a bill on AI in legal practice was pending and a task force was active, and predicted rules within months. The three 2026 bills we could read concern AI in official records and verbatim court records, not attorneys, and show no action after referral; we found no public record of a task force (mncourts.gov was not available to our checks). Neighbor statements were replaced with the Iowa and Illinois facts we had read. (1 page)

Source

Minnesota Legislature, HF 4536 (2026), generative AI in official records (revisor.mn.gov) (read November 3, 2026); Minnesota Legislature, HF 5155 (2026), human oversight in the creation of official verbatim court records (revisor.mn.gov) (read November 3, 2026)

Pages that carry this correction (1)

Michigan state-court page: 'none is pending', 'no task force' and 'no guidance' claims withdrawn

· 1 page

What changed

  • Removed 'none is pending', 'no task force or committee has been publicly announced' and 'no proposed rules, no standing orders, no formal guidance', which we could not verify. The Michigan Court Rules contain no generative-AI text, and a Judicial Council workgroup published a report in October 2024 that is not official Supreme Court policy. We also could not confirm whether the Eastern District's proposed Local Rule 5.1(a)(4) was adopted, and the page now says so. (1 page)

Source

Michigan Court Rules, updated September 2, 2026 (courts.michigan.gov) (read November 3, 2026); Michigan Judicial Council, Generative AI and the Courts Workgroup report, October 2024 (read November 3, 2026)

Pages that carry this correction (1)

Maryland state-court page: Rule 1-311 proposal described from the Rules Committee report; 'imminent' and neighbor claims withdrawn

· 1 page

What changed

  • The page said the proposal specifically addressed AI hallucinations and that formal action was imminent. The report (October 20, 2026) proposes that signing under Rule 1-311(b) certifies the attorney confirmed the existence and authenticity of each cited authority; the text does not mention AI; comment closes November 19, 2026, the Court meets on it on January 5, 2027, and it is not adopted. Unsourced neighbor, D.C. and USPTO statements were removed. (1 page)

Source

Maryland Rules Committee, Two Hundred and Twenty-Ninth Report, October 20, 2026 (mdcourts.gov) (read November 3, 2026)

Pages that carry this correction (1)

Maine state-court page: bar size, forest share and neighbor claims withdrawn; wording tied to the pages read

· 1 page

What changed

  • Removed claims with no source: the size of the bar, the forest share, the aging-bar remark, that Massachusetts and New Hampshire were more active and Vermont had not moved, and the prediction of federal rules. The page now says what the Judicial Branch's index pages show (no AI item), that the Maine State Bar site could not be checked (site unavailable to our checks), and cites Connecticut's Practice Book 4-9. (1 page)

Source

Maine Judicial Branch, administrative orders, rules, proposed rules and committees pages (courts.maine.gov) (read November 3, 2026); Connecticut Law Journal, Practice Book amendments, June 23, 2026 (jud.ct.gov) (read November 3, 2026)

Pages that carry this correction (1)

Louisiana state-court page: 'no commission, no guidance, no legislation' withdrawn; Technology Commission guidelines and Act 250 added

· 1 page

What changed

  • The page said Louisiana had no commission, no guidance and no legislation, gave an attorney count and ranked Louisiana against Texas and Mississippi. The Supreme Court's Technology Commission issued non-binding Generative AI Guidelines on October 3, 2025 for judges and court staff, and Act 250 of 2025 (effective August 1, 2025) changes the Code of Civil Procedure on authenticating false or manipulated evidence. The attorney count and comparisons were removed for lack of a source. (1 page)

Source

Supreme Court of Louisiana Technology Commission, Generative Artificial Intelligence Guidelines, October 3, 2025 (lajudicialcollege.org) (read November 3, 2026); Louisiana Legislature, HB 178, Act 250 of 2025, digest (legis.la.gov) (read November 3, 2026)

Pages that carry this correction (1)

Kentucky state-court page: 'no KBA ethics opinion' and bar-size, ranking claims withdrawn; E-457 and the Whiting sanctions described

· 1 page

What changed

  • The page said the Kentucky Bar Association had published no AI ethics opinion, which contradicted the page's own mention of KBA E-457 (March 15, 2024). It now describes E-457, dates the Whiting opinion (March 13, 2026), states its sanctions and that counsel did not answer the court's question about generative AI, and drops the unsourced attorney count and the ranking against Ohio, Michigan and Tennessee. (1 page)

Source

Kentucky Bar Association, Ethics Opinion KBA E-457, The Ethical Use of Artificial Intelligence in the Practice of Law, issued March 15, 2024 (kybar.org) (read November 3, 2026); U.S. Court of Appeals for the Sixth Circuit, Whiting v. City of Athens, Nos. 24-5918/5919/25-5424, decided March 13, 2026 (ca6.uscourts.gov) (read November 3, 2026)

Pages that carry this correction (1)

Iowa state-court page: 'no task force', 'no formal action' and Minnesota comparison withdrawn

· 1 page

What changed

  • The page said no task force or commission existed and Iowa had taken no formal action. Chapter 27 of the Iowa Court Rules (July 21, 2026, effective January 1, 2027) governs judicial officers' use of AI and creates a judicial branch AI governance committee. Removed the attorney count, the claim that Opinion 512 should be 'treated as binding', the Minnesota comparison and the description of federal judges' standing orders, none with a source. (1 page)

Source

Iowa Judicial Branch, In the Matter of Adopting Chapter 27 of the Iowa Court Rules, Judicial Use of Artificial Intelligence Tools, July 21, 2026 (iowacourts.gov) (read November 3, 2026)

Pages that carry this correction (1)

Idaho state-court page: solo-practice share, 'no pending rulemaking' and neighbor comparisons withdrawn

· 1 page

What changed

  • Removed claims with no source: that roughly 40% of Idaho attorneys practice solo or in small firms, 'no pending rulemaking', the prediction of slower adoption and the Ninth Circuit neighbor comparisons. The bar's blog item is now identified (author, date, what it covers) instead of described as 'blog-style recommendations'. (1 page)

Source

Idaho State Bar, Digital Advocate: How Attorneys Can Think About Artificial Intelligence, by Will Fletcher, October 6, 2025 (isb.idaho.gov) (read November 3, 2026); Idaho Technology Authority, AI & Data Governance Committee (ita.idaho.gov) (read November 3, 2026)

Pages that carry this correction (1)

Hawaii state-court page: 'fake citations triggered the review' and 'deliberate decision' withdrawn

· 1 page

What changed

  • The report does not say that attorney complaints about fake citations prompted the committee, and it does not describe a deliberate decision against disclosure: the committee was established on April 16, 2024 to research AI use and risks, its Ethics Subcommittee proposed no rule changes and recommended a wait-and-see approach, and a proposed Declaration About Use of AI is under consideration by the Hawaii Supreme Court. The page now says that and drops the claims about triggering incidents, enforcement and 'no new rule is needed'. (1 page)

Source

Final Report to the Hawaii Supreme Court, Committee on Artificial Intelligence and the Courts, December 15, 2025 (courts.state.hi.us) (read November 3, 2026)

Pages that carry this correction (1)

Georgia state-court page: Court of Appeals decision re-dated and re-described; 'no committee' claim withdrawn

· 1 page

What changed

  • The decision is dated June 30, 2025, not July 2025. The Court of Appeals vacated a trial court order that relied on non-existent cases (the order appears to have been prepared by one party's attorney) and levied a frivolous motion penalty on that attorney; it said the filings suggest generative AI use, not that the underlying briefing was AI-invented as a finding. Withdrew: 'no committee or task force' (the Judicial Council's Ad Hoc Committee on AI and the Courts was created August 13, 2024 and reported in June 2025; the State Bar has a Special Committee), 'no pending rulemaking', the prediction that rules are coming, and the unsourced federal-court descriptions. (1 page)

Source

Shahid v. Esaam, No. A25A0196, Court of Appeals of Georgia, decided June 30, 2025 (decision text copy, websitedc.s3.amazonaws.com) (read November 3, 2026); Judicial Council of Georgia, Ad Hoc Committee on Artificial Intelligence and the Courts, Findings and Recommendations, June 2025 (copy hosted at nowhabersham.com) (read November 3, 2026)

Pages that carry this correction (1)

Connecticut state-court page: federal-judge AI orders no longer asserted

· 1 page

What changed

  • Removed the statement that District of Connecticut judges have issued AI orders: the Practice Book notice covers state court only and no federal order was read. (1 page)

Source

Connecticut Law Journal, Practice Book amendments, June 23, 2026 (jud.ct.gov) (read November 3, 2026)

Pages that carry this correction (1)

Arkansas state-court page: 'no guidance, a task force is coming' replaced by Administrative Order No. 25

· 1 page

What changed

  • The page said Arkansas had no guidance and only a task force. The Supreme Court adopted Administrative Order No. 25 (Artificial Intelligence) on December 11, 2025 and published proposed Rules of Professional Conduct amendments on AI in June 2025, on a recommendation of the Arkansas Bar Association's AI Task Force. Removed the unsourced bar size, the count of states, the neighbor comparisons (which contradicted our own corrected Oklahoma and Missouri pages) and the prediction. We read the decisions in a mirror because the court's site was blocked to our reader. (1 page)

Source

Arkansas Supreme Court, In re Adoption of Administrative Order No. 25, Artificial Intelligence, 2025 Ark. 220, December 11, 2025 (case-law mirror hallapproved.com; the court's opinions site was not available to our reader) (read November 3, 2026); Arkansas Supreme Court, In re Amendments to the Arkansas Rules of Professional Conduct, 2025 Ark. 115, June 5, 2025 (case-law mirror) (read November 3, 2026)

Pages that carry this correction (1)

Arizona state-court page: 'October 2024' policy date, 12-18 month prediction and Maricopa judge orders withdrawn

· 1 page

What changed

  • Withdrawn without a source: the month of the policy (we read the January 2025 amending order, not the original), the prediction that attorney rules follow court-personnel policies within 12-18 months, the claim that Maricopa County judges have de facto disclosure orders, a comparison to 'Texas's Fifth Circuit', and the statement that the State Bar had issued no AI guidance (it has published guidance). Added the August 27, 2026 Supreme Court order denying a Rule 135 petition. (1 page)

Source

Arizona Supreme Court, Administrative Order 2025-24 amending ACJA section 1-509 (azcourts.gov), January 29, 2025 (read November 3, 2026); State Bar of Arizona, Guidance for the Use of Generative Artificial Intelligence in the Practice of Law in Arizona (azbar.org) (read November 3, 2026)

Pages that carry this correction (1)

Alaska state-court page: bar size, 'ahead of most states' and no-committee claims withdrawn

· 1 page

What changed

  • Removed claims with no source: that Alaska is 'ahead of most states', the size of the bar, that the Alaska Supreme Court has formed no committee and no judge has an AI order, and that Washington and California have more developed frameworks. The page now gives the opinion's adoption dates and says that the opinion tells lawyers to confirm whether court rules or orders require disclosure. (1 page)

Source

Alaska Bar Association, Ethics Opinion 2025-1, Generative Artificial Intelligence & The Practice of Law (alaskabar.org), approved April 3, 2025, adopted April 23, 2025 (read November 3, 2026)

Pages that carry this correction (1)

Southern District of Florida page: no AI order found on judge pages, not a partial framework

· 1 page

What changed

  • The page said the district had a partial AI framework, that individual judges had addressed AI in their courtrooms and that disclosure was effectively required. None of the 27 district-judge pages we read posts an AI order and the local-rule amendments effective December 1, 2026 contain no AI provision. The page now says there is no district-wide rule and no AI order found, and it no longer calls the district a circuit benchmark. (1 page)

Source

Southern District of Florida, 27 district and senior judge pages and linked PDFs (flsd.uscourts.gov) (read November 3, 2026); Southern District of Florida, Administrative Order 2026-94, Amendments to the Local Rules, effective December 1, 2026 (flsd.uscourts.gov)

Pages that carry this correction (1)

Best AI for personal injury lawyers: 'best' and 'highest ROI' claims about EvenUp reworded as our judgment

· 1 page

What changed

  • The page said EvenUp is the best AI tool for personal injury lawyers, that it delivers the highest ROI and has the most direct impact on revenue. We found no independent head-to-head test and no measured ROI, and EvenUp publishes no price. The page now says EvenUp is our first pick for firms with a demand-package bottleneck, as our judgment, and lists what each plaintiff-focused vendor publishes. (1 page)

Source

EvenUp (evenuplaw.com, read November 3, 2026): vendor claims; no price published, Schedule a Call; Supio (supio.com, read November 3, 2026): vendor-published description; Book A Demo

Pages that carry this correction (1)

NotebookLM 'Plus' monthly price replaced by Google AI Pro at $19.99/month on the NotebookLM for legal page

· 1 page

What changed

  • The page gave NotebookLM Plus its own monthly price while also saying Plus is not sold separately. It now says expanded NotebookLM access comes with Google AI Pro, which Google lists at $19.99 a month. The CoCounsel price on the page (from $340.85 a month for in-house teams, Thomson Reuters page) was re-read on November 3, 2026 and did not change. (1 page)

Source

Google, Gemini subscriptions (gemini.google, read November 3, 2026): Google AI Pro $19.99 / month

Pages that carry this correction (1)

Claude Mythos 5.1 access programs and Opus 4.7 cybersecurity-safeguard details re-read against Anthropic's announcements on 4 pages

· 4 pages

What changed

  • Removed statements that Opus 4.7's cybersecurity safeguards are classifier infrastructure in the base model, run at the model layer, are more reliable than Opus 4.6's enforcement and flag named categories (phishing pretexts, surveillance tooling, exploit generation). Anthropic's Opus 4.7 release notes say only that the model ships with safeguards that automatically detect and block requests indicating prohibited or high-risk cybersecurity uses, and that it is the first model to carry them. The pages now say that and that the categories, location and reliability are not published. (2 pages)
  • Removed claims that the safeguards separate analysis of a disclosed exploit from exploit generation, behave identically on every surface, create a retained log entry and show the user a refusal, a Freshfields reference, and a measured reduction in jailbreak risk. Anthropic's release notes state none of this; the pages now mark them as not verified and tell firms to test their own prompts and ask Anthropic or the platform. (2 pages)
  • The statement that insurance carriers will start, or are beginning, to ask about default cybersecurity safeguards is now labeled as our forecast, not a reported trend; no carrier source was found. (2 pages)
  • Mythos 5.1 access now names the two programs Anthropic lists. The Cyber Verification Program currently covers certain Opus- and Sonnet-class models and will add Mythos-class models 'in the near future', so cyber access to Mythos 5.1 is not stated as available today; the Life Sciences Verification Program has enrolled its first participants. Fable 5.1 may be used to identify vulnerabilities, but penetration testing, exploit generation and binary-based scanning are still redirected to Opus models. 'Same model family' became 'same underlying model', as Anthropic words it. (2 pages)

Source

Anthropic, Introducing Claude Fable 5.1 and Claude Mythos 5.1 (anthropic.com, September 2026, read November 3, 2026): same model, Mythos 5.1 only through trusted access programs; the Cyber Verification Program will include Mythos-class models in the near future; the Life Sciences Verification Program has enrolled its first participants; Fable 5.1 safeguards still redirect penetration testing, exploit generation and binary-based vulnerability scanning to Opus models; Anthropic, Developing Enterprise Frontier Safeguards with our customers (anthropic.com, September 1, 2026, read November 3, 2026); Anthropic, Introducing Claude Opus 4.7 (anthropic.com, April 16, 2026, read November 3, 2026): safeguards that automatically detect and block requests indicating prohibited or high-risk cybersecurity uses; no categories, location in the stack, log or per-surface behavior stated

Pages that carry this correction (4)

GPT-4, GPT-4o and ChatGPT-4 named as what ChatGPT plans, chatbots or vision tools run, and the 'consumer Claude and GPT-4o match or exceed enterprise legal AI' verdict, replaced on 9 pages

· 9 pages

What changed

  • Removed GPT-4, GPT-4o and ChatGPT-4 as what ChatGPT's Free and Plus plans run, as the vision model for medical records, as the model behind legal chatbots and custom builds, and as the example of a current LLM or of a tool name courts ask for. OpenAI's models page lists GPT-6 models as current and does not list GPT-4 or GPT-4o, and we have not verified which model each plan runs, so the pages now say OpenAI's models, ChatGPT, or that we have not verified it. (7 pages)
  • Removed the verdict that consumer Claude and GPT-4o match or exceed enterprise legal AI (solo-lawyer guide), and the generalization that GPT-4 performs better than GPT-3.5 but still fabricates citations in niche areas (ChatGPT risks guide). Neither had a current source: the solo-lawyer page now says we have not benchmarked consumer assistants against enterprise legal AI, and the risks page limits the Stanford 2024 figures to the 2024 models. The Stanford 2024 GPT-4 figure itself and Allen & Overy's 2023 GPT-4 deployment are dated history and stay. (2 pages)

Source

OpenAI, API models (developers.openai.com, read November 3, 2026): current models GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna; GPT-4o and GPT-4 are not listed; the model page for gpt-5.4-codex returned 404; Google, Gemini API pricing (ai.google.dev, read November 3, 2026): gemini-3.1-pro-preview is listed at $2 / $12 per 1M tokens for prompts up to 200k tokens ($4 / $18 above); Anthropic, models overview (platform.claude.com, read November 3, 2026): current Claude models list a 1M-token context window

Pages that carry this correction (9)

Middle District of Pennsylvania page: four judges with AI sections, not none

· 1 page

What changed

  • The page said no Middle District judge had issued an AI order and that the district had remained silent. The court's own judge pages show four that address AI: Judge Mehalchick, Magistrate Judge Latella and Magistrate Judge Camoni require a Certificate of Use of Generative AI, and Magistrate Judge Caraballo warns parties to verify AI output. There is still no district-wide rule. The page now says so. (1 page)

Source

Judge Karoline Mehalchick, Use of Generative AI (pamd.uscourts.gov) (read November 3, 2026); Magistrate Judge Leo Latella, Use of Generative AI (pamd.uscourts.gov) (read November 3, 2026); Magistrate Judge Sean Camoni, Use of Generative AI (pamd.uscourts.gov) (read November 3, 2026); Magistrate Judge Phillip Caraballo, Use of Generative AI (pamd.uscourts.gov) (read November 3, 2026)

Pages that carry this correction (1)

30 judge pages no longer title themselves 'AI Standing Order for Attorneys (2026)' when the judge's posted orders contain no AI provision

· 30 pages

What changed

  • The titles of these judge pages, and on nine of them the search description, presented an AI standing order from the judge, with disclosure, verification and sanctions terms. The body of each page had already been re-read against the court's own posted orders on November 2-3, 2026 and found no AI provision (for Judge Scholer the rule is N.D. Tex. Local Rule 7.2(f), for Judge Hanks S.D. Tex. General Order 2025-04, both district-wide). The titles now say what the re-read shows. Each page has fewer than 5 Google clicks in 28 days and the new titles keep the judge's name and the words AI and order. (30 pages)

Source

Judge Donato page on the N.D. Cal. site (read November 2-3, 2026); Judge Breyer general standing order (rev. July 16, 2025) (read November 2-3, 2026); Judge Caproni page on the S.D.N.Y. site (read November 2-3, 2026); Judge Koeltl Individual Rules (March 2025) (read November 2-3, 2026); Judge Kelly page on the D.D.C. site (read November 2-3, 2026); Chief Judge Connolly page on the D. Del. site (read November 2-3, 2026); Judge Noreika page on the D. Del. site (read November 2-3, 2026); Judge Kennelly page on the N.D. Ill. site (read November 2-3, 2026); Judge Burroughs page on the D. Mass. site (read November 2-3, 2026); Judge Liman page on the S.D.N.Y. site (read November 2-3, 2026); Judge Preska page on the S.D.N.Y. site (read November 2-3, 2026); Judge Nuffer chambers page on the D. Utah site (read November 2-3, 2026); S.D. Texas General Order 2025-04, Use of Generative AI in Court Filings (district-wide; read November 3, 2026)

Pages that carry this correction (30)

Judge pages (Kim, Noreika, Lee, Torres, Cannon, Kelly, Beeler, Walker, Birotte, Howell): wording re-read against the courts' own pages and the Federal Judicial Center

· 9 pages

What changed

  • The Kim page said her orders were not reviewed for discovery provisions and added a Silicon Valley caseload remark without a source. Both of her civil and discovery orders (October 7, 2025 and October 23, 2024) were read in the court PDFs: neither mentions AI, and the discovery section is now described as written. The filler is removed. (1 page)
  • The Noreika page gave a count of Delaware judges that the court's own page contradicts, summarized Chief Judge Connolly's orders and added Delaware corporate-law filler, none with a primary source. The court page (no AI order listed) and the Federal Judicial Center record were read; confirmation and commission dates are corrected and the unsourced statements are removed. (1 page)
  • The Lee page compared her order with other N.D. Cal. judges' orders it had not read, offered an opinion on using Claude and described her courtroom as a node of AI litigation. Section H of her Civil Standing Order (August 31, 2026) was re-read and matches; the comparisons, the opinion and the filler are removed. (1 page)
  • The Torres page described her docket and compared her order with those of other S.D.N.Y. judges, with no primary source. Section VII of her Individual Practices in Civil Cases (revised November 5, 2025) was re-read and matches; the docket description and comparisons are removed. (1 page)
  • The Cannon page said she became one of the most recognized federal judges in the country, that every procedural choice she makes draws national attention, and that no AI provision exists in her posted procedures. The court site could not be read, so the page no longer says either way; her dates now follow the Federal Judicial Center and the filler is removed. (1 page)
  • The Kelly page repeated press reports (a Senate inquiry, the Heppner ruling, a D.C. Circuit ruling in Thaler) and compared S.D.N.Y. judges as if read. The court page and Federal Judicial Center record could not be read on November 3, 2026, so what only the press supports is removed from the page. (1 page)
  • The Walker page narrated an earlier mistake, described the court's local rules as a 2015 edition and disclaimed a docket description. His Federal Judicial Center record and the Florida administrative orders were re-read; the Garcia ruling is attributed to press coverage and the stale statements are removed. (1 page)
  • The Birotte page listed the contents and dates of orders that his court page does not show, and compared them with Judges Scarsi and Hwang. His Federal Judicial Center record was read; the page now says it cannot confirm whether he has an AI order and the unsupported statements are removed. (1 page)
  • The Howell page listed her on the court's Information Technology Committee, predicted that "I prompted the AI" would not be enough and said a Copyright Office guidance note was unavailable. Her Federal Judicial Center record and the D.D.C. pages show she served on that committee from October 2013 to March 2016 and now sits on Criminal Law; the prediction and the note are removed. (1 page)

Source

Federal Judicial Center, Biographical Directory of Article III Federal Judges, read November 3, 2026, for appointment and senior status dates; N.D. Cal. judges' standing orders (Kim, Lee, Beeler), read November 3, 2026; S.D.N.Y. Individual Practices in Civil Cases (Torres), read November 3, 2026; D. Del. judge page (Noreika), read November 3, 2026; D.D.C. senior-judge and standing-orders pages (Howell), read November 3, 2026; Florida administrative orders AOSC26-12 and 11th Circuit AO 26-18 (Walker), read November 3, 2026

Pages that carry this correction (9)

Judge pages (Thompson, Hwang, Scarsi, Otake, Young, Chutkan, Fitzgerald, Rosenberg, Casper, Durkin): wording re-read against the courts' own orders and the Federal Judicial Center

· 10 pages

What changed

  • The Thompson page called her court 'one of the most tech-forward' and dated her criminal and patent orders as revised in August. Her orders (read November 3, 2026) are dated September 3, 2026 (criminal pretrial/trial and patent) and August 22, 2025 (criminal); the Noland decision is a state decision and California Rule 10.430 governs state courts, not federal filings. Prestige and docket claims were removed and the practical steps are labelled as our suggestions. (1 page)
  • The Hwang page said her civil order had been reissued with the same paragraph 5. We read the current civil order (October 21, 2026) and criminal order (September 25, 2026) but not an earlier version, so that statement was removed; her sanctions section (H) was added and the district-rule statement now says what we checked. (1 page)
  • The Scarsi page called his order one of the clearest in the district and said paragraph 15 was unchanged. Paragraph 15 was re-read (separate declaration naming the AI-generated part; sanctions warning); the earlier version was not read, so 'unchanged' and the ranking were removed, and his career now follows the Federal Judicial Center. (1 page)
  • The Otake page said we did not know what General Order 23-1 requires. The order, a scanned PDF read by OCR on November 3, 2026, requires a 'Reliance on Unverified Source' statement filed with a brief generated by an unverified source; it is dated November 14, 2023 (her procedures say November 13). Docket composition and the count of active judges had no primary source and were removed. (1 page)
  • The Young page described the E.D. Va. pace as a fact and built a national-security section on an unsourced docket. His order of July 16, 2026 is a Rule 7(E) order with no AI provision; 'Rocket Docket' is now a press nickname without statistics, and the advice is labelled as ours. (1 page)
  • The Chutkan page's meta description promised specific AI disclosure requirements. Her court page, Standing Orders page and the September 2026 Local Rules contain no AI provision. The description now says so; prestige adjectives and an unsourced D.C. Courts task force date were removed, and the Trump and DOGE matters are attributed to press coverage. (1 page)
  • The Fitzgerald page described his docket as among the busiest and gave AI error frequencies. His procedures site and five PDFs contain no AI text; the volume and error-rate claims had no primary source and were removed, and the class action, fee and employment sections are labelled as our suggestions. (1 page)
  • The Rosenberg page said she chaired the federal rules committee until 2025. Her chambers page says chair since 2022, and the June to September 2026 roster lists a different chair. The Eleventh Circuit item now reflects AO 26-18 rescinding AO 26-15, and the 17th Circuit item AO 2026-19-Gen; the earlier orders are attributed to The Florida Bar News. (1 page)
  • The Casper page described the Boston legal market, a patent docket and a prior OCDETF role as fact. Her court page, six PDFs and the Local Rules (effective September 8, 2026) contain no AI text; none of those claims appears in the court page or the Federal Judicial Center, so they were removed and her career now follows the FJC. (1 page)
  • The Durkin page described the Chicago bar and his docket as fact. His case procedures contain no AI text. The Illinois Supreme Court policy (effective January 1, 2025) was read and now states that disclosure of AI use should not be required in a pleading; market, volume and docket claims had no primary source and were removed. (1 page)

Source

Federal Judicial Center, Biographical Directory of Article III Federal Judges (judges.csv, read November 3, 2026): nominations, confirmations and commissions; N.D. Cal., Judge Thompson's standing orders and the court's 'Using AI tools in your case' guide, read November 3, 2026; C.D. Cal., Judges Hwang, Scarsi and Fitzgerald published procedures (JpsApi), read November 3, 2026; D. Haw., General Order 23-1 and Judge Otake's procedures, read November 3, 2026; E.D. Va., Judge Young's order of July 16, 2026 and Chief Judge Lauck's order of September 1, 2026, read November 3, 2026; D.D.C., Judge Chutkan's court page, Standing Orders and Local Rules (September 2026), read November 3, 2026; S.D. Fla. and D. Mass. chambers pages and Local Rules; N.D. Ill. Judge Durkin's case procedures and the Illinois Supreme Court AI policy, read November 3, 2026

Pages that carry this correction (10)

Judge pages (Baylson, Johnston, Lin, Scholer, Cooper, Ellis, Hanks, Martinez-Olguin, Curiel): wording re-read against the courts' own orders and the Federal Judicial Center

· 9 pages

What changed

  • The Baylson page compared his order with other judges' yes/no certifications and described Philadelphia's law firms and docket without a source. The order of June 6, 2023 was read in the court's PDF; those two statements are removed. (1 page)
  • The Johnston page said he took a different approach from most federal judges and summarized critics' arguments from an article we had not read. His order text on the court site was read; the comparison and the critics' arguments are removed, Starr's order is described as the court now describes it, and the Illinois policy is no longer characterized. (1 page)
  • The Lin page said the order was not separate and did not say it also binds self-represented litigants. Her September 4, 2026 order was re-read: it applies to unrepresented litigants too, and the class-action reference is the proposed settlement administrator's declaration. Her civil trial and criminal orders were not read, and the page now says so. (1 page)
  • The Scholer page described an 'unusual background' and a commercial docket without a source. Her court page (no AI order) and N.D. Tex. Local Rule 7.2(f) were read; career dates now follow the Federal Judicial Center and the unsourced claims are removed. (1 page)
  • The Cooper page described his docket as among the country's most consequential and quoted Rule 11(b) as covering 'every factual allegation'. The D.D.C. standing orders and local rules (September 2026) contain no AI provision and Rule 11(b) speaks of contentions; docket claims are removed and Federal Judicial Center dates added. (1 page)
  • The Ellis page attributed 'national press attention' and a speed temptation to the case mix without a source. Her court page and the N.D. Ill. local rules have no AI provision; Fuentes' civil order (March 3, 2026) was read by OCR. Press-only items are now labelled as press coverage. (1 page)
  • The Hanks meta description said his order sets disclosure, filing format and sanctions, and the page described his docket mix. General Order 2025-04 was read: it cautions about verifying AI-assisted filings and requires no disclosure, certificate or format. Meta description, FAQ and unsourced docket and district claims are corrected. (1 page)
  • The Martinez-Olguin page speculated about Silicon Valley and presented two of our own suggested steps as if they came from her order. Section H paragraph 4 (revised August 26, 2026) was re-read; the speculation is removed and the suggested steps are labelled as ours. (1 page)
  • The Curiel page presented Trump University, a settlement, a cartel threat and a Marshals relocation as fact and applied California's Rule 10.430 to federal filings. The AI sections of his rules were read; press-only items are attributed, Rule 10.430 is described as governing state courts, and Federal Judicial Center dates are added. (1 page)

Source

Federal Judicial Center, Biographical Directory of Article III Federal Judges (judges.csv, read November 3, 2026): appointment and senior status dates; N.D. Cal. judges' standing orders (Lin, Martinez-Olguin, Curiel's S.D. Cal. rules), read November 3, 2026; N.D. Ill. judges' pages and local rules (Johnston, Ellis, Fuentes' civil order), read November 3, 2026; N.D. Tex. Local Civil Rule 7.2(f) and Judge Scholer's court page, read November 3, 2026; S.D. Tex. General Order 2025-04 (Judge Hanks' court page), read November 3, 2026; D.D.C. standing orders and local rules (Judge Cooper), read November 3, 2026

Pages that carry this correction (9)

Judge pages (Starr, Fuentes, Novak, Kronstadt, Woods, Donato, Connolly, Singhal, O'Connor, Kacsmaryk): wording re-read against the courts' own orders and the Federal Judicial Center

· 10 pages

What changed

  • The Starr page gave the date, the 'first' label, the tools named and his reasoning for the 2023 order as fact. The order is no longer posted, so its text is unavailable and those points now say 'according to press reports' or are removed. His court page and N.D. Tex. Local Rule 7.2(f) (effective September 1, 2026) were read and are described as written. (1 page)
  • The Fuentes page called the order 'a notably balanced policy' and its meta description promised disclosure and sanction rules. The AI section was re-read in the order itself: it has no disclosure requirement and no sanctions text. The Illinois Supreme Court policy it points to could not be read, so its summary is marked unverified. (1 page)
  • The Novak page described a former federal prosecutor on a fast docket without a source (the Federal Judicial Center lists no prior career and the court's biography page was unavailable). His October 1, 2026 order concerns Local Rule 7(E) and has no AI text; the AI rule in the district is Chief Judge Lauck's September 1, 2026 order, which requires identifying AI use and certifying citations. (1 page)
  • The Kronstadt page described an entertainment docket, the district's size and a list of AI risks in patent cases without sources. Three published C.D. Cal. documents of his contain no AI text. Dates now follow the Federal Judicial Center and the claim about judges with AI orders is stated as a floor. (1 page)
  • The Woods page called S.D.N.Y. 'ground zero' for AI litigation and credited him with a first AI privilege ruling and committee leadership, none with a source. His Individual Rules (revised March 19, 2025) have no AI text; Judge Broderick's rules (revised October 29, 2025) do require disclosure and certification. (1 page)
  • The Donato page said he was a Shearman & Sterling partner and a federal prosecutor; the Federal Judicial Center lists a deputy city attorney career. His seven posted orders contain no AI text. California Rule of Court 10.430 is restated: effective September 1, 2025, courts adopt a policy by December 15, 2025, covering their own use rather than filings. (1 page)
  • The Connolly page claimed a disproportionate share of corporate, patent and bankruptcy litigation, 'most important venue' and 'groundbreaking' status without sources. The Fortune 500 share is restated from corp.delaware.gov, and the April 18, 2022 citizenship order was added. The orders we read contain no AI text; scheduling and special master orders were not read. (1 page)
  • The Singhal page claimed a busy docket, an international-litigation hub and maritime risks without sources. FJC dates were added. No AI text appears on his page, in the Local Rules effective December 1, 2025 or in administrative order 2026-94. (1 page)
  • The O'Connor page described a culture of disclosure that does not exist as a rule of his. His page has no AI text; the district rule is Local Rule 7.2(f). The Supreme Court reversed the preventive-care ruling on June 27, 2025 (Kennedy v. Braidwood), now stated. Superlatives and predictions were removed. (2 pages)
  • The Kacsmaryk page called the April 7, 2023 mifepristone ruling a nationwide injunction; it was a stay under section 705 of the APA (Doc. 137). The oath quotation was corrected, the 'only active judge in Amarillo' claim and the 'most consequential federal judges' superlative were removed as unsourced, and Local Rule 7.2(f) was added. (2 pages)

Source

Federal Judicial Center, Biographical Directory of Article III Federal Judges (judges.csv, read November 3, 2026): appointment and senior status dates; N.D. Tex., Local Civil Rule 7.2(f) (CIVRULES.pdf, footer 'Effective September 1, 2026') and Judge Starr's court page, read November 3, 2026; N.D. Ill., Judge Fuentes' standing order (generative AI section), read November 3, 2026; E.D. Va., Judge Novak's order of October 1, 2026 and Chief Judge Lauck's order of September 1, 2026, read November 3, 2026; C.D. Cal., judges' published procedures (Judge Kronstadt), read November 3, 2026; S.D.N.Y. Individual Rules (Judge Woods), N.D. Cal. standing orders (Judge Donato) and D. Del. orders (Judge Connolly), read November 3, 2026

Pages that carry this correction (10)

Judge pages (Coleman, Altman, E. Chen, Fischer, Wright, Pittman, Furman, Stein, Brinkema, Padin): wording re-read against the courts' own orders and the Federal Judicial Center

· 10 pages

What changed

  • The Coleman page did not say that the Seventh Circuit vacated the approval of the Clearview AI class settlement on July 13, 2026 (the order approving it is dated May 12, 2025). The page now states both dates; the 'most significant case' and 'algorithmic bias' framing had no source and is removed. (1 page)
  • The Altman page called him one of the most active judges of the Southern District of Florida and described his docket; none of that had a source and it is removed. Appointment dates now come from the Federal Judicial Center, and the page records that his chambers page posts no AI standing order. (1 page)
  • The Chen page gave a 'revised 2/6/2025' date for his civil standing order; the file is dated 12-1-2022, and none of his six posted orders mentions AI. The 'first Asian American' and '161 years' claims, the Senate vote count, the Korematsu, Uber and Oscar Grant claims and a second Ninth Circuit appeal in Nosal had no primary source and are removed (one 2016 appeal is verified). Senior status is dated May 17, 2022. (1 page)
  • The Fischer page described landmark entertainment cases without a source (removed). We read nine of her posted documents and found no AI text; senior status is dated May 1, 2024. The $31,100 sanction is attributed to press reports of a special master's order, not to the order itself; the $10,000 Noland sanction matches the opinion. (1 page)
  • The Wright page said he has no generative AI order. His Judge's Procedures on the C.D. Cal. site (section VII.A) require an AI certificate on the last page of every memorandum of points and authorities and of pretrial, trial and posttrial briefs, and ask filers to disclose generative AI use. The earlier read had only opened the attached PDFs. The Prenda sanction is $81,319.72 with referrals to the state bars, the U.S. Attorney and the IRS; the page no longer says FBI or 'six-figure'. The C.D. Cal. district page repeated the 'no order' sentence and is corrected. (1 page)
  • The Pittman page described a high-profile and regulatory docket without a source; those sections are removed. Local Civil Rule 7.2(f) and Local Criminal Rule 47.2(e) were re-read; his 2019 appointment is attributed to the Federal Judicial Center. (1 page)
  • The Furman page now covers all four sets of practices (civil, pro se, criminal and hearings/trials; the last two were revised October 26, 2026), none with AI text. United States v. Heppner is attributed to the law-firm commentary because the docket could not be read. (1 page)
  • The Stein page ranked the Times case as the best known; that had no source and is removed. The page now cites the MDL (No. 25-md-3143) and the order of January 5, 2026, and the cross-references to Judges Broderick, Engelmayer, Subramanian and Castel carry links and read dates. (1 page)
  • The Brinkema page said the Google trial began in 2025; it began in September 2024 and the liability ruling is dated April 17, 2025 (Department of Justice release). The remedies description is attributed to press, the Moussaoui entry follows the Fourth Circuit's September 13, 2004 decision, and the Snowden and 'biggest in history' claims are removed. (1 page)
  • The Padin page gave a revision date for her Judicial Preferences that the page itself does not show; the date is removed and the read date is November 3, 2026. Her biography now follows the Federal Judicial Center. (1 page)

Source

Federal Judicial Center, Biographical Directory of Article III Federal Judges (judges.csv, read November 3, 2026): appointment and senior status dates; C.D. Cal., Judge Otis Wright's Judge's Procedures, section VII.A (Artificial Intelligence), read November 3, 2026; Seventh Circuit, Weissman v. Clearview AI (No. 25-1673, decided July 13, 2026), read November 3, 2026; N.D. Cal., Judge Edward Chen's standing orders (six PDFs) and the court's AI page, read November 3, 2026; S.D.N.Y., Judge Furman's practices page and the Advisory Committee on Evidence Rules report (May 2026), read November 3, 2026; N.D. Tex., Local Civil Rule 7.2(f) and Local Criminal Rule 47.2(e), read November 3, 2026

Pages that carry this correction (10)

Judge Lucy Koh page no longer titles itself 'AI Standing Order for Attorneys (2026)': the Ninth Circuit's rules contain no AI order of hers

· 1 page

What changed

  • The title and search description of this page presented an AI standing order from Judge Koh, with disclosure, verification and sanctions terms. The Ninth Circuit's rules handbook says its rules do not regulate how a filing is produced, including with generative AI, and that the signature attests the signer reviewed it; its General Orders contain no AI text, and we found no AI order of Judge Koh's. The title now says no AI order was found, the description says what the circuit's rules do say, and the page body states the verdict and links the primary. The page has 2 Google clicks in 28 days. (1 page)

Source

Ninth Circuit Rules handbook, advisory committee note to Circuit Rule 32-1 (new December 1, 2025) (read November 3, 2026); Ninth Circuit General Orders effective October 27, 2026 (read November 3, 2026)

Pages that carry this correction (1)

Conflict checks, Harvey ROI and AI audit pages: hour, price and settlement figures with no source removed

· 3 pages

What changed

  • The Harvey ROI list said lawyers need 20-40 hours to become proficient with Agent Builder and that deployment takes 3-6 months. Neither figure had a source and we found no Harvey publication that gives either, so the page now says that and tells the reader to time a pilot and ask for a written timeline. (1 page)
  • The conflict-checks page said AI conflict tools cost $200-1,000 a month, less than an hour of partner time, and that conflict-related malpractice claims average $150,000-300,000 in settlements. No source supported any of these figures, so the page now gives none and tells the reader to get written quotes. (1 page)
  • The AI audit guide said a small firm can self-audit in 1-2 days of partner time and gave audit consultant fees of $10,000-25,000 and $25,000-75,000. No source supported those figures, so the answer now gives no figure and tells the reader to count their own hours and ask for a written quote. (1 page)

Source

Harvey, Agents page, no training time or deployment time published (read November 3, 2026); Lawmatics pricing page, pricing set per practice with a demo request and no price list (read November 3, 2026)

Pages that carry this correction (3)

Freshfields and Anthropic: the claim that Freshfields is 'the public reference' for Anthropic deals replaced with what the firms announced

· 10 pages

What changed

  • Ten pages called Freshfields 'the public reference' for firms with active Anthropic deals. No source says that. Freshfields' own press release of April 23, 2026 announces a multi-year collaboration with Anthropic: Claude access for 5,700 employees, a commitment to firmwide deployment across 33 offices and a co-development program. The pages now say that and nothing more; whether other firms have comparable deals is not something we have verified. (10 pages)

Source

Freshfields, 'Freshfields and Anthropic team up to co-build AI legal workflows, deploying Claude across the firm globally' (press release, April 23, 2026; read November 3, 2026)

Pages that carry this correction (10)

Claude Fable 5.1 guides: plan availability, retention wording and cloud-route rows re-read against Anthropic, Microsoft and AWS documentation

· 3 pages

What changed

  • The Fable 5.1 deployment checklist listed the product for Pro, Max, Team and Enterprise without the limits, and said Fable requires 30-day retention by default. Anthropic's pricing page lists Fable on Pro through usage credits, on Max and Team premium seats within 50% of weekly limits, and on Enterprise, not on Free. Anthropic's Covered Models article says prompts and completions are retained for at least 30 days on every surface and that zero data retention is not available on the standard route. Both statements now say that. (1 page)
  • The cloud procurement guide said Google's Model Garden lists Fable 5.1 for Vertex AI Studio and that Microsoft lists it with Entra ID authentication. Google's console listing could not be re-read without a session, so that row now cites Anthropic's naming (Google Cloud Agent Platform) and says the listing is unverified. Microsoft Learn (dated September 21, 2026) lists Fable as Preview on Anthropic infrastructure and names Entra ID only for Mythos, while Anthropic calls Fable generally available on Foundry; the row states both. (1 page)
  • The retention guide said prompts and outputs are retained for 30 days. Anthropic says at least 30 days, then deleted unless under a safety investigation or legal hold; the guide now says at least. (1 page)

Source

Anthropic, Introducing Claude Fable 5.1 and Claude Mythos 5.1 (read November 3, 2026): availability, Google Cloud Agent Platform, EFS and ZDR; Anthropic Help Center, Covered Models (read November 3, 2026): 30-day retention, surfaces, ZDR for eligible customers; Anthropic, Claude pricing (claude.com/pricing, read November 3, 2026): Fable by plan; Microsoft Learn, Claude models in Microsoft Foundry (page dated September 21, 2026, read November 3, 2026)

Pages that carry this correction (3)

Claude Fable 5.1 guides: Google Cloud row read from Google's documentation, retention wording and plan attribution re-checked

· 5 pages

What changed

  • The cloud procurement guide said Google's own listing was not re-read and the Vertex AI Studio detail was unverified. Google's public documentation page for Fable 5.1 is readable without a session: it lists claude-fable-5-1 as generally available since September 1, 2026, with 1M input and 128K output tokens and a retirement date no sooner than March 1, 2027, and it states prompts and responses are stored for up to 30 days, where Anthropic's Covered Models article says at least 30 days. The row now says that and asks buyers which retention applies. Microsoft Learn was re-read and still lists Fable as Preview on Anthropic infrastructure. (1 page)
  • Four guides still said Fable prompts are deleted after 30 days, retained for 30 days, subject to a 30-day default requirement, or kept 30 days for zero-data-retention organizations. Anthropic's Covered Models article says a 30-day minimum by default: retained at least 30 days, then deleted unless under a safety investigation or legal requirement, with zero data retention not available on the standard route. The guides now say that. (4 pages)
  • Anthropic's pricing page lists plan access for Fable without separating Fable 5 from Fable 5.1, so the deployment checklist now says so instead of implying the limits are specific to 5.1. (1 page)

Source

Google Cloud, Anthropic's Claude Fable 5.1 on Google Cloud (page updated October 30, 2026, read November 3, 2026); Anthropic Help Center, Covered Models (read November 3, 2026): 30-day minimum retention, surfaces, ZDR; Anthropic, Claude pricing (claude.com/pricing, read November 3, 2026): Fable by plan; Microsoft Learn, Claude models in Microsoft Foundry (page dated September 21, 2026, read November 3, 2026)

Pages that carry this correction (5)

Fable 5.1 guides (legal benchmark, 5 to 5.1 migration, text watermark): every figure and attribution re-read against Anthropic's system card, model docs, release post, watermark page and Covered Models page

· 2 pages

What changed

  • The migration guide's retention row said the covered-model rule applies unless an exception is approved, and a second row said serving infrastructure can change separately from weights. Anthropic's Covered Models page lists Fable 5.1 (designated August 31, 2026) and describes retention of at least 30 days on every surface plus a limited-time ZDR option for eligible customers; the row now says that. The release post says Fable 5.1 and Mythos 5.1 are the same model with different safeguard levels; the serving-infrastructure sentence had no source and is replaced. The 1M-token context, 128K output and $0.25 per million cache-read figures were re-read in the model docs and are unchanged. (1 page)
  • The watermark guide said heavy rewriting can remove the signal. Anthropic's page says light editing probably will not remove the watermark completely and that a complete rewrite where every word is replaced will; the guide now says that. The Aug. 14 announcement, Sept. 1 update and private-preview detector were re-read and are unchanged. (1 page)

Source

Claude Fable 5.1 and Claude Mythos 5.1 System Card, section 8.15.2 Legal Agent Benchmark (read November 3, 2026); Anthropic, Claude Fable 5.1 model overview (context, output, pricing; read November 3, 2026); Anthropic, Introducing Claude Fable 5.1 and Claude Mythos 5.1 (read November 3, 2026); Anthropic, Claude text watermark (published Aug 14, 2026, updated Sep 1, 2026; read November 3, 2026); Anthropic Help Center, Covered Models (Fable 5.1 designated August 31, 2026; read November 3, 2026)

Pages that carry this correction (2)

District of the Virgin Islands page: no D.V.I. AI order exists

· 1 page

What changed

  • The page described a D.V.I. AI disclosure order, a certification requirement and a pro hac vice compliance duty. We found no such order from the court or its own judges (court website and linked PDFs read November 3, 2026), so those statements are gone. One visiting judge's posted policies (Judge Kearney, E.D. Pa.) warn that AI hallucinations can draw sanctions; the page now says that instead. (1 page)

Source

Judge Mark A. Kearney, Policies and Procedures for actions in the District Court of the Virgin Islands, February 2026, sections I.H and I.L (vid.uscourts.gov)

Pages that carry this correction (1)

Microsoft Copilot for law firms: the $30 price is confirmed, the unsourced April 15 feature list and SKU details are removed

· 1 page

What changed

  • Removed the claim that Microsoft's April 15, 2026 release added contract comparison, missing-provision detection, Outlook client email drafting, Teams deposition summaries and Excel billable-hours analysis. The April 15 report we cite lists three Word capabilities: Track Changes with word-level precision, contextual comments and a table of contents, available at launch only through the Frontier program on the Office Insiders Beta Channel. The page now says that. The $30 per user per month figure stays: Microsoft's enterprise pricing page lists the Microsoft 365 Copilot add-on at $30.00 per user per month, paid yearly, with a qualifying Microsoft 365 plan required (read November 3, 2026). (1 page)
  • Removed the list of qualifying Microsoft 365 plans (E3, E5, F3) and base-license prices for E3, E5, Business Standard and Business Premium, and the 300-user limit on the Business add-on. Microsoft's pricing page says a qualifying Microsoft 365 plan is required but does not list them, so the page now says to confirm them with a Microsoft representative. (1 page)

Source

Microsoft, Copilot pricing, Enterprise plans tab: Microsoft 365 Copilot (add-on) $30.00 user/month, paid yearly, qualifying Microsoft 365 plan required (microsoft.com, read November 3, 2026); Artificial Lawyer, April 15, 2026 (quoting Microsoft): track changes, contextual comments and table of contents in Word; Frontier program on the Office Insiders Beta Channel (read November 3, 2026); Microsoft, Copilot pricing, Business plans tab: Business Premium with Copilot $32.00, Business Standard with Copilot $23.50, Copilot Business add-on $18.00 (originally $21.00), paid yearly (read November 3, 2026)

Pages that carry this correction (1)

Claude individual plans re-read against Anthropic's pricing page; Astra availability marked as not re-verified

· 4 pages

What changed

  • Three pages described Claude Max as starting 'from' a price. Anthropic's pricing page now lists Max as a single tier and a separate Ultra plan above it; the pages now name both plans with their prices, dated November 3, 2026. (3 pages)
  • The cyber-model comparison described OpenAI Astra as a limited rollout, not yet generally available, based on a September 3, 2026 reading. OpenAI's page refused our request when we tried to re-read it on November 3, 2026, so the page now says Astra's current status is not verified. (1 page)

Source

Anthropic, Claude pricing (claude.com/pricing; read November 3, 2026): Pro $17 a month annual or $20 monthly, Max $100 a month, Ultra $200 a month

Pages that carry this correction (4)

Circuit Courts pages (14): claims re-read against the courts' own sites, rules and orders; unsourced or wrong statements removed or marked as unconfirmed

· 14 pages

What changed

  • Corrected November 3, 2026: An earlier version said DOJ policy requires disclosure of AI-assisted work product, that DOJ and agency AI policies create a de facto standard private practitioners must match, and that agency frameworks universally require human oversight. We could not find a published DOJ or agency rule saying so; DOJ's April 2026 annual report lists its Interim AI Policy as pending. We also removed statements about D.D.C. judges' standing orders and a local rules committee that we could not confirm. The D.C. Circuit and D.D.C. rules were rechecked and contain no AI provision. (1 page)
  • We re-read this page against primary sources. The Eighth Circuit has seven states, not eight (28 U.S.C. 41). The District of Nebraska has Local Civil Rule 7.1(d)(2), effective December 1, 2025, which requires parties who use generative AI to verify their filings; the page had said every district in the circuit lacked formal AI rules and called the circuit uniformly quiet. We removed the statement that Missouri's 20th Judicial Circuit is the only court in the circuit with an AI disclosure mandate and the claims that it is a nationwide outlier or a de facto federal standard. The 20th Circuit's rule is now described only according to secondary summaries. We also removed unsourced statements about the circuit's opinions and about Minnesota judges' standing orders. (1 page)
  • November 3, 2026 Removed the statements that the Eleventh Circuit has more active AI disclosure requirements than any other circuit and that compliance is mandatory (the November 2 removal had missed the Bottom Line); that Georgia's vacatur makes AI disclosure effectively mandatory across the circuit; that its judicial council, annual conference or district courts have addressed AI disclosure; that the Middle District of Florida issued district-wide guidance; that the Northern District of Georgia and Alabama's federal judges have issued AI sanctions or orders; and that courts in the circuit have ordered attorneys to produce documentation of their AI workflows. We found no source for them. We also corrected the Georgia ruling: in Shahid v. Esaam (No. A25A0196, June 30, 2025) the Georgia Court of Appeals vacated one trial court order that relied on two fictitious cases and imposed a $2,500 penalty on the opposing lawyer; it made no finding on who inserted the citations and does not require disclosure of AI use. The page now cites Akerlund v. Atlas Air (No. 24-11033, July 10, 2026, for publication), which states a duty to verify AI output and refers counsel to the circuit's disciplinary committee, but creates no disclosure rule. We have not surveyed the Florida, Georgia or Alabama federal districts' orders. (1 page)
  • Corrected November 3, 2026: An earlier version attributed to Finnegan an analysis identifying at least five tribunal-specific AI rules; we could not locate that analysis and removed it. We also corrected the statement that PTAB judges require parties to disclose AI use (the USPTO's April 2024 guidance says there is no general obligation to disclose), removed the reference to covered business method proceedings (that program ended for new petitions on September 16, 2020), and removed claims about AI rules at the Court of Federal Claims and the ITC that we could not confirm. What we did find: USPTO guidance and individual AI orders from four Court of International Trade judges. We also removed unsourced damages figures. (1 page)
  • November 3, 2026 The Fifth Circuit page said the court 'voted' and that its 'judicial council' decided on candor and competence grounds, described a Fifth Circuit standing order requiring AI certification in appellate briefs, and cited a 'Proposed Rule 7.07' that would impose national AI disclosure. The court's own notice says it decided not to adopt a rule governing the use of generative AI in briefs and reminds counsel of their duties under FRAP 46(b)(1)(B); a three-judge AI Subcommittee recommended against the rule. The court has no AI certification order, and the national proposal is Rule of Evidence 707 on machine-generated evidence, not a disclosure rule. We also removed unsupported statements about the Southern, Eastern and Western Districts of Texas and Louisiana, and named the Mississippi judge (Henry T. Wingate) from the Senate Judiciary release. (1 page)
  • The page said the Massachusetts SJC interim guidelines address AI use in court filings and set the practical standard for attorneys, and that the Second Circuit produced Mata v. Avianca and has issued AI guidance. The SJC interim guidelines cover generative AI use by judges and court personnel, not attorneys' filings; Mata v. Avianca is a Southern District of New York decision; the Second Circuit has no AI order. We also removed statements about First Circuit judges' expectations, informal warnings, a First Circuit advisory and District of Massachusetts standing orders that we could not source. (1 page)
  • November 3, 2026 The Fourth Circuit page said Virginia 'just passed' HB1642 and described it as an AI transparency law covering legal services, with a disclosure duty for AI that affects client outcomes and statutory penalties. HB1642 is a 2025 act (Acts 2025, c. 637, Va. Code § 19.2-11.14) about human decision-makers in criminal-justice decisions that use AI-based tools; it has no disclosure duty for law firms and no penalty provision. We also removed statements about federal agency AI policies, Maryland and other districts, and the Fourth Circuit's administrative office that we could not confirm in court sources. (1 page)
  • We re-read this page against primary sources. California Rule of Court 10.430 requires each California court to adopt a policy on generative AI use by its own staff and judicial officers. It does not require attorneys to disclose AI use in filings, define material contribution or set certification language; the page had presented it as the national benchmark for attorney disclosure. Those statements are removed. Hawaii's Supreme Court committee called Rule 11 adequate for now but proposed an AI declaration for filers, and the federal District of Hawaii has General Order 23-1, so the earlier line that Hawaii declared existing rules sufficient is rewritten. We removed the statement that King County Superior Court and several Western District of Washington judges require AI disclosure, which we could not source, and the statement that the Ninth Circuit had issued no direct ruling, and added its June 3, 2026 order in LNU v. Blanche. (1 page)
  • Corrected November 3, 2026: An earlier version said no justice has publicly addressed AI; Chief Justice Roberts discussed AI and hallucinated citations in his 2023 Year-End Report on the Federal Judiciary. We also removed statements about what Supreme Court practitioners say or do with AI, a count of amicus filings per term, the claim that the Court has no formal sanctions framework, and claims about a professional norm of disclosing AI use, none of which we could source. The statement that the Court has no AI rule was rechecked against the Rules of the Supreme Court effective March 16, 2026 and the Clerk's 2026 guides, and still holds. (1 page)
  • The page described United States v. Heppner (S.D.N.Y., Judge Rakoff) as a ruling on communications between attorneys and AI tools. The ruling concerned a represented defendant's own conversations with a public AI tool, which were held not privileged. We also corrected Judge Starr's court (Northern District of Texas) and removed statements we could not source about an Eastern District of New York chief judge's guidance, EDNY standing orders, an expected EDNY policy, SDNY show-cause orders, a magistrate judge's quotation and Second Circuit judges' remarks. (1 page)
  • We re-read this page against primary sources. The Seventh Circuit's published rules contain no generative AI provision. The Northern District of Illinois orders we read (Judge Johnston, Judge Coleman) require compliance with Rule 11(b) and Rule 26(g) and do not ask for a disclosure statement; the page had called Illinois judges the most vocal opponents of AI disclosure in the country and credited unnamed prominent judges with statements we could not source. We also removed claims that this view influenced other districts and that the Southern and Central Districts of Illinois followed the Northern District. The reference to Whiting v. Athens as an AI sanctions case is removed, because the Sixth Circuit did not find that AI was used; the Seventh Circuit's own case is Dec v. Mullin. (1 page)
  • November 3, 2026 The Sixth Circuit page said several judges in the Northern and Southern Districts of Ohio require AI disclosure, that at least two Northern District of Ohio judges imposed informal sanctions over poorly executed AI-assisted briefing, and that Michigan judges focus on AI in e-discovery protocols. We could not verify those statements in court sources and removed them. The page now cites only Judge James L. Graham's October 21, 2025 Southern District of Ohio order. We rechecked the Whiting figures against the opinion and found them correct. (1 page)
  • November 3, 2026 Removed the statements that the Tenth Circuit has issued no guidance on AI in filings, that its appellate bench has not addressed attorney AI use, that no committee studies the issue, that no federal district in its six states has AI rules, that Colorado's act reaches AI-assisted case evaluation or litigation strategy, and the unsourced statements about Utah's sandbox and about Colorado's act shifting expectations across the circuit. The page now says what the court records show: in Moore v. City of Del City (No. 25-6002, unpublished, December 3, 2025) the Tenth Circuit dismissed a pro se appeal in the alternative for misuse of generative AI and ordered that litigant to disclose AI use in future filings; in Picon-Diaz v. Bondi (No. 25-9530, February 13, 2026) it warned counsel about an apparently fabricated citation; and Judge Nina Y. Wang of the District of Colorado has required an AI certification in every filing since December 1, 2025. We did not survey every district in the circuit. (1 page)
  • The page said Pennsylvania's interim AI policy requires attorneys to ensure accuracy and recommends disclosure in filings, and that Delaware's Court of Chancery is developing AI guidelines. The Pennsylvania policy governs judicial officers and court personnel, and we found no court-wide Chancery AI rule, only case-specific orders. We also removed the statement that the ROSS opinion determines which AI tools lawyers may legally use; the opinion decides copyright and fair use for a non-generative tool. (1 page)

Source

D.C. Circuit Rules and Handbook of Practice (Sept 2026 edition), read November 3, 2026; D.D.C. Local Rules, September 2026, read November 3, 2026; DOJ Chief Data Officer Annual Report FY2025 (April 2026), read November 3, 2026; EOIR PM 25-40, read November 3, 2026; 28 U.S.C. 41, read November 3, 2026; Nebraska Civil Rules (NECivR) 7.1(d)(2), effective December 1, 2025, read November 3, 2026; D. Minn. GenAI FAQ, October 1, 2025, read November 3, 2026; Eighth Circuit Local Rules, June 2024, read November 3, 2026; Georgia Court of Appeals, Shahid v. Esaam, No. A25A0196, June 30, 2025 (opinion PDF copy), read November 3, 2026; Eleventh Circuit, Akerlund v. Atlas Air, No. 24-11033, July 10, 2026, read November 3, 2026; Florida Supreme Court, AOSC26-12, May 28, 2026, read November 3, 2026; USPTO guidance on AI tools, Federal Register 2024-07629, read November 3, 2026; USPTO covered business method program page, read November 3, 2026; USPTO revised inventorship guidance (Nov 2025), read November 3, 2026; Court of International Trade, Judge Vaden order, read November 3, 2026; Federal Circuit Rules of Practice, read November 3, 2026; Fifth Circuit, Court Action on Proposed AI Rule (ca5.uscourts.gov, read November 3, 2026); Fifth Circuit, Court Decision on Proposed Rule (ca5.uscourts.gov, read November 3, 2026); Fletcher v. Experian, No. 25-20086 (5th Cir. Feb. 18, 2026) (ca5.uscourts.gov, read November 3, 2026); Report of the Advisory Committee on Evidence Rules, Dec. 1, 2025 (uscourts.gov, read November 3, 2026); Senate Judiciary Committee, Grassley release of judges' responses, Oct. 23, 2025 (judiciary.senate.gov, read November 3, 2026); Mass.gov, Interim Guidelines for Use of Generative AI (scope as shown in the page result; direct fetch returned 403), read November 3, 2026; U.S. Court of Appeals for the First Circuit, home page, read November 3, 2026; U.S. Court of Appeals for the Second Circuit, home page and local rule notices, read November 3, 2026; Va. Code § 19.2-11.14, Use of artificial intelligence-based tool (law.lis.virginia.gov, read November 3, 2026); HB1642, 2025 Regular Session (lis.virginia.gov, read November 3, 2026); Fourth Circuit website (ca4.uscourts.gov, read November 3, 2026); California Rules of Court, rule 10.430, read November 3, 2026; Hawaii Committee on AI and the Courts, Final Report, December 15, 2025, read November 3, 2026; D. Haw. 2026 General Civil Procedures (General Order 23-1), read November 3, 2026; LNU v. Blanche, No. 24-4790 (9th Cir. June 3, 2026), read November 3, 2026; Rules of the Supreme Court (effective March 16, 2026), read November 3, 2026; Supreme Court Amicus Guide 2026, read November 3, 2026; 2023 Year-End Report on the Federal Judiciary, read November 3, 2026; Seventh Circuit rules (FRAP, Circuit Rules, Operating Procedures), read November 3, 2026; N.D. Ill. Judge Johnston case procedures, read November 3, 2026; N.D. Ill. Judge Coleman standing orders, read November 3, 2026; Dec v. Mullin, No. 25-2417 (7th Cir. Mar. 30, 2026), read November 3, 2026; S.D. Ohio, Standing Order on the Use of Generative Artificial Intelligence, Judge James L. Graham, Oct. 21, 2025 (ohsd.uscourts.gov, read November 3, 2026); Whiting v. City of Athens, Nos. 24-5918/5919/25-5424 (6th Cir. Mar. 13, 2026), opinion PDF on Justia, read November 3, 2026; Tenth Circuit, Moore v. City of Del City, No. 25-6002, order and judgment of December 3, 2025, read November 3, 2026; Tenth Circuit, Picon-Diaz v. Bondi, No. 25-9530, February 13, 2026, read November 3, 2026; D. Colo., Judge Nina Y. Wang, standing order on generative AI (effective December 1, 2025), read November 3, 2026; Unified Judicial System of Pennsylvania, Interim Policy on Generative AI (applies to judicial officers and court personnel), read November 3, 2026; Third Circuit, Thomson Reuters v. ROSS Intelligence, No. 25-2153, read November 3, 2026; E.D. Pa., Standing Order re generative AI, Judge Pratter (May 3, 2024), read November 3, 2026

Pages that carry this correction (14)

Central District of California page: at least 18 judges with an AI provision, not three

· 1 page

What changed

  • The page said at least three judges had posted AI orders. Reading all 61 judge entries on the court's procedures site, at least 18 have an AI provision in their orders or procedures and at least 14 require a declaration, certificate or disclosure. The page now says so and names them. (1 page)

Source

Central District of California, Judges' Procedures app (apps.cacd.uscourts.gov/Jps), all 61 judge entries (read November 3, 2026); Judge Mark C. Scarsi, Civil Case Standing Order, October 2026, paragraph 15 (apps.cacd.uscourts.gov)

Pages that carry this correction (1)

'Verified citations' and 'guarantees real citations' no longer stated as features of Westlaw AI, CoCounsel and Lexis+ AI

· 56 pages

What changed

  • These pages described Westlaw AI, CoCounsel and Lexis+ AI as giving verified citations, verified case law or guaranteed real citations. Thomson Reuters describes verified Westlaw content and LexisNexis describes Shepard's tools for checking citations, but neither promises that every citation an AI answer gives is real or correct, and a Stanford study found both vendors' research tools hallucinate between 17% and 33% of the time. The pages now say the answers are grounded in the vendor's database and link to sources that still have to be checked. (56 pages)

Source

Magesh et al., 'Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools' (Stanford RegLab and HAI, arxiv 2405.20362, read November 2, 2026): Westlaw AI-Assisted Research and Lexis+ AI hallucinated between 17% and 33% of the time, after vendors had described their tools as eliminating, avoiding or guaranteeing no hallucinated citations; Thomson Reuters, CoCounsel Legal product page (read November 2, 2026): describes 'verified Westlaw content'; no promise that citations are real or verified; LexisNexis, Lexis+ with Protégé product page (read November 2, 2026): 'Shepard’s Verify checks citations in AI responses, documents, and drafts'; Shepard's Citations to confirm status and treatment

Pages that carry this correction (56)

Claude, ChatGPT, Harvey, Legora, Luminance, Lexis, Westlaw, Everlaw, Relativity and Nuix: 'X wins' and 'better for' verdicts without a published benchmark reworded or removed

· 10 pages

What changed

  • Removed the statements that Claude's legal writing is 'consistently superior to every other general-purpose AI', 'cleaner than ChatGPT', better 'than any competitor' (Perplexity page) and that ChatGPT Plus is 'weaker than Claude on complex legal reasoning'. No published benchmark ranks these tools on legal writing; the Vals Legal AI Report evaluated no general-purpose model and no drafting task. The pages now say the ranking is our judgment, not a measured result. (4 pages)
  • Removed 'Harvey wins in the US. Legora wins in Europe' (headline, summary and bottom line), 'the only option that actually works', Harvey's listed weaknesses (civil law, languages, training corpus), 'deeper UK common law training', Legora's '16 European countries' and its EU-native-infrastructure claim. No independent benchmark compares them, Harvey does not publish its training data, and Legora's own pages give no country count or hosting region. The page now names no winner and points to a pilot. (1 page)
  • Removed 'DISCO is the better choice' for smaller matters and 'Relativity still wins on extreme-scale matters ... Everlaw usually wins on simplicity-adjusted cost'. We found no independent benchmark of scale, extensibility or cost per matter; the page now says to compare quotes on your own matter profile. (1 page)
  • Replaced the verdict that Lexis wins for natural-language searchers and Westlaw for Boolean researchers with the Stanford study's measured figures (about 17% versus 33% hallucinated answers, 2024 product versions, CoCounsel and Harvey not tested). No source shows which product suits which search style. (1 page)
  • Removed 'Luminance handles contract tasks better' and 'Harvey is the better first investment for litigation-dominant firms / Luminance for transaction-dominant firms'. Neither vendor appears in a published benchmark; the page now says to pilot both on your highest-volume practice area. (1 page)
  • Replaced 'For standard litigation document review: Relativity's aiR is the better tool' with what Relativity says aiR does (relativity.com, read November 2, 2026); Nuix's home page describes no equivalent review feature and no independent benchmark compares them. (1 page)
  • Removed 'Harvey's legal reasoning is measurably superior, its training data is deeper, and its outputs require less verification'. No measurement compares Harvey with Claude, Harvey does not publish its training data, and Stanford found 17% to 33% hallucination in the research tools it tested, so the page says to verify either way. (1 page)

Source

Vals Legal AI Report (updated February 27, 2025; read November 2, 2026, evaluated CoCounsel, Vincent AI, Harvey Assistant and Oliver on seven tasks, no Claude, ChatGPT, Legora or Luminance, and no legal drafting task; Magesh et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, Stanford (arXiv 2405.20362; read November 2, 2026, Lexis+ AI about 17% and Westlaw AI-Assisted Research about 33% hallucinated answers; Legora security page (read November 2, 2026, SOC 2 Type II, ISO 27001, GDPR; enterprise customers control where data is stored, no hosting region, market count or language list; Relativity aiR (read November 2, 2026, case strategy, review and Assist, no comparison with Nuix

Pages that carry this correction (10)

Vendor capability claims without a primary (Spellbook in Word, Luminance leases and due diligence, Harvey hiring-tool and deposition analysis) attributed to the vendor's pages or removed

· 3 pages

What changed

  • The real estate page said Spellbook drafts and redlines purchase agreements and leases directly inside Word, flags non-standard lease clauses in seconds, and that Luminance reviews hundreds of leases at once, was built for due diligence and learns each jurisdiction. Spellbook's own page supports Word and redlining but not real estate or any timing; Luminance's pages describe comparing up to 10,000 contracts against playbooks but not leases, due diligence or jurisdictions. The page now quotes what each vendor says, with the date we read it, and drops the rest. (1 page)
  • The employment guide said Harvey can analyze a client's AI hiring tools against NYC Local Law 144 and the Illinois AI Video Interview Act and maps compliance obligations across every jurisdiction, and that Harvey can analyze deposition transcripts. Harvey's platform page describes regulatory-change tracking, legal research and bulk document analysis, and mentions neither hiring tools nor depositions. Those statements are replaced by what the page says. (1 page)
  • The Harvey versus Westlaw AI guide said Harvey is measurably better at complex legal analysis and produces analysis CoCounsel cannot match. No benchmark supports it. The page now quotes Harvey's own description of its research tool and says we have not benchmarked it against CoCounsel. (1 page)

Source

Spellbook, Review feature page (read November 2, 2026): 'Add precise redlines to your contracts... right in Microsoft Word'; the pages read do not mention real estate, leases or purchase agreements; Luminance, repository product page (read November 2, 2026): 'Compare up to 10,000 contracts at once against your playbooks'; the pages read do not mention leases, rent escalations, break clauses, due diligence or jurisdictions; Harvey, platform page (read November 2, 2026): Horizon Scanning, Knowledge ('complex legal, regulatory, and tax questions across domains') and Vault ('bulk analyze legal documents'); no mention of hiring tools, NYC Local Law 144 or deposition transcripts

Pages that carry this correction (3)

Response-speed and intake figures removed, because no primary source measures them for law firms

· 5 pages

What changed

  • The page said studies consistently show the first firm to respond to a prospective client gets the case 60-70% of the time. No study we could open measures that share, so the page now says speed of response matters and that we have no measured share. (1 page)
  • The page said responding to a lead within 5 minutes increases conversion by 400% compared with 30 minutes. We found no primary source for that figure for law firms, so the page now says we have no sourced conversion figure. (1 page)
  • The page said the average law firm takes 2-3 days to respond to an inquiry and that a potential client contacts 3-5 firms and hires the first to respond, citing unnamed legal industry studies. None was named or found, so both figures are removed and the page says we have no measured figure. (1 page)
  • The page said automated intake is the difference between converting 20% and 40% of leads. No source measures that lift, so the page now says we have no measured conversion lift. (1 page)
  • The page attributed to the Clio Legal Trends Report a 42% share of potential clients who hire another attorney after failing to reach a firm by phone. No such figure appears in the Clio material we found, so the share is removed. (1 page)

Source

Harvard Business Review, The Short Life of Online Sales Leads (Oldroyd, McElheran, Elkington, March 2011), the study these speed claims are usually traced to; it studies online sales leads, not law firms, and its body was behind a paywall on November 2, 2026, so no multiplier is quoted

Pages that carry this correction (5)

Shared tool-card prices (Relativity per-GB hosting, ChatGPT Team, CoCounsel seat estimate) without a primary replaced by the vendor pages read November 2, 2026

· 19 pages

What changed

  • This page said Relativity is priced per GB for hosting and AI features under enterprise contracts. Relativity's pricing page says the price is set by committed data volume with AI included and no per-user fees, and publishes no per-GB or dollar rate, so the card now says that. (5 pages)
  • This page listed ChatGPT plan prices, including Team at $25 per user per month. OpenAI's pricing page returned 403 when we read it on November 2, 2026, so we could not confirm any current price and the card now says so instead of quoting one. (7 pages)
  • This page listed Claude Team at $25 per user per month without the billing term. Anthropic's pricing page lists $25 billed monthly and $20 billed annually for a standard seat, so the card now gives both. (6 pages)
  • This page described CoCounsel pricing as seat-based. Thomson Reuters' page we read does not publish a seat rate, so the card now says firm plans are quoted and cites the in-house price from its page read October 25, 2026. (1 page)

Source

Relativity, aiR pricing page (relativity.com/platform/pricing, read November 2, 2026): price is set by committed data volume, AI included, no per-user fees, pay as you go or a one- or three-year term; the only call to action is Request Pricing. It names no dollar or per-GB rate.; Anthropic, Claude pricing (claude.com/pricing, read November 2, 2026): Team standard seat $25 a month billed monthly or $20 billed annually, 2 to 150 seats; Pro $20 monthly or $17 annual; OpenAI, ChatGPT pricing (chatgpt.com/pricing, tried November 2, 2026): the page answered 403 Forbidden to our read, so no current plan price could be confirmed and none is published here

Pages that carry this correction (19)

Superlatives about Harvey ('most powerful legal AI platform', 'best legal AI platform', 'gold standard', 'clear market leader', 'strongest') replaced with a description or an attribution

· 11 pages

What changed

  • These pages called Harvey the most powerful, the best, the most capable or the gold standard of legal AI, the clear market leader, or the strongest on data protection or brief writing. Harvey's own site makes no such ranking and cites no independent benchmark, and we found none. The pages now describe Harvey as an enterprise legal AI platform sold by quote, attribute its security statements to its security page, and say where we found no market-share figure. (11 pages)

Source

Harvey, home page (read November 2, 2026), where the only 'most powerful' is 'The most powerful Harvey yet' about Harvey II, and Harvey describes itself as 'AI software for legal and professional services' and cites no ranking, analyst award or independent benchmark; Harvey, security page (read November 2, 2026), which says customer data is logically separated between customers and that inputs, outputs and uploaded documents are not used to train underlying models unless the customer opts in

Pages that carry this correction (11)

'Faster than a human', 'most powerful AI', 'unmatched' and 'best-in-class' claims without a primary reworded on 8 pages

· 8 pages

What changed

  • The best AI for solo practitioners page said AI handles discovery responses, documents, research memos and contract review faster than a human. We found no measured comparison, so the page now says to time it on your own work. (1 page)
  • The Relativity vs DISCO vs Everlaw page said Relativity has the most powerful AI, Everlaw the most usable and DISCO the fastest, called Relativity's extensibility unmatched and said Everlaw offers 90% of its capability; the Relativity pricing page called it the most powerful e-discovery platform on the market. None of the three vendors' sites publishes such a ranking, no independent head-to-head test exists that we found, and the 90% figure had no source. The pages now give our reading of each feature set, say plainly that it is not a benchmark, and tell readers to compare total cost of ownership on their own quote. (2 pages)
  • Five pages called one vendor's feature unmatched or each tool best-in-class (Westlaw's AI and database, Clio's flexibility, Bloomberg Law's compliance alerts, Luminance's contract review). No primary or independent comparison supports those rankings, so the pages now describe the feature as that tool's main strength and no longer rank it against competitors. (5 pages)

Source

Relativity, home page (relativity.com, read November 2, 2026): no ranking, benchmark or 'most powerful' claim; DISCO, home page (csdisco.com, read November 2, 2026): carries a G2 'Fastest implementation' badge, which is about setup, not AI processing speed; Everlaw, home page (everlaw.com, read November 2, 2026): no ranking or benchmark of AI speed or power; Luminance, home page (luminance.com, read November 2, 2026): no claim that its review is unmatched

Pages that carry this correction (8)

Claude Mythos 'most powerful AI model' replaced with Anthropic's own wording

· 1 page

What changed

  • The Claude Mythos page called Mythos Preview Anthropic's most powerful AI model. Anthropic's own announcement describes it as a general-purpose, unreleased frontier model and does not rank it, so the page now uses that wording, says it is Anthropic's description, and no longer ranks it. (1 page)

Source

Anthropic, Project Glasswing announcement (April 7, 2026, read November 2, 2026), which calls Claude Mythos Preview a general-purpose, unreleased frontier model and does not call it the most powerful; Anthropic Frontier Red Team, Mythos Preview write-up (read November 2, 2026), which calls it a new general-purpose language model

Pages that carry this correction (1)

Four unsourced superlatives reworded (Legalweek networking, Judge Boasberg's experience, incident training, converted-skeptic testimonial)

· 4 pages

What changed

  • Four pages used unsourced superlatives: the legal AI conferences guide said Legalweek's networking is unmatched, the Judge Boasberg page said his FISC service gave him unparalleled experience, the AI incident response plan called real incidents the most powerful training material, and the AI implementation failures guide called a converted skeptic the most powerful testimonial. None had a source, so each now states a plainer, supportable version (the conference line is labeled our editorial read). (4 pages)

Source

Legalweek event page (read November 2, 2026), which gives no attendee survey or networking comparison; for the other three statements we found no source at all

Pages that carry this correction (4)

Six pages that kept retired claims: an unsourced count of federal courts, Florida circuit orders, a Northern District of California claim, Colorado SB 205 in a description and FAQ, Briefpoint listed for deposition summaries, and Copilot recovery and cost totals

· 7 pages

What changed

  • Removed a count of federal district courts that require AI disclosure. No primary gives that count and we could not find the RedGrave LLP 2025 update the page cites; the page now says some federal judges and state courts require disclosure and that we have not counted them. (1 page)
  • Replaced a statement that some judges in the Northern District of California require clear identification of AI-assisted documents while others allow AI use. The court's own guide for self-represented litigants says some judges have disclosure requirements, and Judge Chen's published orders do not mention AI. (1 page)
  • Removed the claim that the Eleventh Circuit has more AI disclosure activity than almost any other circuit and the search-result description that said Florida has three active judicial circuit orders. AOSC26-12 bars Florida courts from imposing their own AI disclosure or certification requirements. (1 page)
  • Replaced the Colorado page's description and two FAQ answers that called SB 205 the first impact-assessment law and gave February 2026 enforcement. SB 24-205 was repealed and reenacted by SB 26-189, a narrower law that applies to consequential decisions made on or after January 1, 2027; the answers now say we have not mapped it to law firm uses. (1 page)
  • Removed four statements that Briefpoint is built for deposition summaries and beats general AI at them, including a 30-45 minute turnaround. Briefpoint's own site describes written-discovery drafting, and we have not timed any deposition tool. (1 page)
  • Removed productivity-recovery ranges, a 2-6 week breakeven, configuration hours and combined licensing totals from the Copilot ROI comparison and procurement guide. They rested on hours nobody measured; the pages now give the $30 per user per month add-on price and say to time a pilot. (2 pages)

Source

Briefpoint, home page (briefpoint.ai, read November 2, 2026): drafting and responding to written discovery (interrogatory answers, RFP responses), no deposition summaries; Florida Supreme Court, AOSC26-12, May 28, 2026 (read October 25, 2026); Judge Chen, standing orders and guidelines (cand.uscourts.gov, read October 23, 2026): none mentions AI; Colorado General Assembly, SB26-189 (signed May 14, 2026; applies to consequential decisions made on or after January 1, 2027), read October 28, 2026; Microsoft 365 Copilot pricing, read October 31, 2026: $30.00 per user per month, paid yearly

Pages that carry this correction (7)

Open re-reads closed: Eleventh Circuit 'Alabama will follow', 'epicenter of AI litigation' for the Northern District of California, Copilot training and developer hours and year-two 'pure margin', and Colorado SB 26-189 in the page body

· 6 pages

What changed

  • Removed the claims that Alabama will follow the circuit's disclosure practices, that the circuit's culture is clearly pro-disclosure, and that the Eleventh Circuit is the most pro-disclosure, most rigorous or among the strictest in the country. We found no source for any of them, and the page already said we have not ranked the circuit. For Alabama filings the page now says we have not surveyed its federal judges' orders and to check the assigned judge's order. (1 page)
  • Removed the statements that the Northern District of California is the epicenter of AI litigation and handles more AI-related cases than any other federal district, and that its judges apply a higher standard to AI-related attorney conduct. No dataset counts AI-related cases by district. On the Judge Chen page we now give what our own lists show: four of the eight AI training-data copyright cases on our list are in the district (a hand-checked list, not a census), and 27 U.S. decisions on AI-fabricated citations come from it, fewer than the Southern District of New York (55), the Eastern District of Michigan (38), the Western District of Washington (34) and the Central District of California (33). The Judge Alsup page no longer says epicenter. (2 pages)
  • Removed training hours (4-8 per attorney, 8-15 per paralegal, 4-6 for a small firm), a 40-200 hour developer estimate for the Cowork legal plugin, and the claim that year-two Copilot productivity recovery is 'pure margin'. Nobody measured them; the pages say we have not measured them and keep the $30 per user per month add-on price. (2 pages)
  • The Colorado page's body now describes SB 26-189 (signed May 14, 2026): a narrower automated decision-making law with notice, human-review and developer-documentation duties, applying to consequential decisions made on or after January 1, 2027. We replaced the sentence that said it defines high-risk AI systems and sets a threshold that pulls law firms into scope, which was the 2024 act's wording and is not what the bill page we read says; the page now says we have not mapped the act's definitions to law firm uses. (1 page)

Source

Colorado General Assembly, SB26-189 (signed May 14, 2026; applies to consequential decisions made on or after January 1, 2027; 60 days' notice and a chance to cure before any action before January 1, 2030), read October 28, 2026; AI Vortex, AI training data copyright lawsuits: eight hand-checked US cases, four in the Northern District of California (docket and orders re-read October 31, 2026); AI Vortex sanctions tracker, decisions on AI-fabricated citations by court, read November 2, 2026: 27 U.S. decisions from the Northern District of California, against 55 from the Southern District of New York

Pages that carry this correction (6)

Fourteen pages: unsourced 'X% of the capability at Y% of the cost' ratios removed

· 14 pages

What changed

  • Removed four claims that Claude, ChatGPT or a general-purpose stack delivers a fixed share (70-80%) of a specialist tool's value for 5-15% of its cost (contract review, under-$100 stack, CLM for small firms, criminal-defense research). No vendor publishes such a ratio and we found no independent benchmark. The pages now cite the published Claude Pro price with the read date, or say that Juro and Harvey publish no comparable price and that we have no benchmark. (4 pages)
  • Removed claims that AI patent-search tools and Claude deliver 80-90% of a professional search firm's or specialist tool's output at 0.5-20% of the cost. No benchmark or published price supports a ratio; the pages now say we have no independent benchmark and, for Claude, cite the published Pro price with the read date. (2 pages)
  • Removed claims that Claude or ChatGPT cover 70-80% of Harvey's or CoCounsel's work at 5-10% of the cost. Harvey and CoCounsel do not publish comparable prices and no benchmark exists; the pages now cite the published Claude Team seat price with the read date and say how much of the work is covered is untested. (4 pages)
  • Removed the claim that a mid-market firm gets 70% of the Freshfields-Anthropic operational benefit at 5% of the cost (two pages). No source gives either figure; the pages now say part of the benefit is capturable and that we have no benchmark of how much. (2 pages)
  • Removed the claims that Logikcull covers 80% of cases at 30% of Relativity's cost and that vLex Vincent costs 25 to 30 percent of Westlaw Precision with CoCounsel. Neither price ratio is published or benchmarked; the pages now point readers to current vendor quotes. (2 pages)

Source

Anthropic, Claude pricing page (read November 2, 2026): Pro $20 per month on monthly billing ($17 per month billed annually); Team standard seat $25 per seat per month on monthly billing ($20 billed annually)

Pages that carry this correction (14)

Statements that we ran our own tests removed from six pages: no test, log or dataset exists behind them

· 6 pages

What changed

  • The Gemini vs Claude legal writing page said we ran five legal writing tasks on both tools and that Claude was stronger in four, and its direct answer said the pick came from an earlier test. We have no record of any such test, so the page now says we have not run one, lists the five tasks as a way to test and no longer ranks the two on writing quality. (1 page)
  • The best AI for bankruptcy lawyers page said Claude handled the Bankruptcy Code better than ChatGPT in our testing. We did not run that test and found no published comparison, so the page now says so and keeps the instruction to verify every citation. (1 page)
  • The Spellbook vs Claude page said we used both tools on the same contract tasks and printed minutes per task, a 20-40% speed advantage and roughly $2,000 a month in recovered billing. No such timings were taken. The timings and the figures derived from them are removed; the page now lists three tasks to time on your own documents, and the verdict no longer calls Spellbook the better contract tool. (1 page)
  • The Claude Mythos for litigators page said the author tested Claude for 30 days on active litigation tasks with strong results. There is no record of that test, so the sentence is replaced by a description of a test a litigator can run on a closed matter. (1 page)
  • The Claude Opus 4.7 guide told an employment-law test of Texas and California non-competes in the first person, with a result. There is no record of the run, so the page now presents the question as a test to run, says we have not run it and drops the conclusion drawn from the anecdote. (1 page)
  • The Claude vs Perplexity page said its split between drafting and research held across every legal use case we had tested. We ran no such tests; the page now calls it an editorial framework. (1 page)

Source

AI Vortex site repository and Lagos working files searched November 2, 2026; no test record, log, prompt set or dataset was found for any of these statements

Pages that carry this correction (6)

Unsourced price ranges ($5,000-15,000) and 'armies of associates' claims removed or rewritten as quote-based wording on 22 pages

· 22 pages

What changed

  • This page gave a $5,000-15,000 figure for cannabis compliance retainers. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says there is no published figure and to get a written quote. (1 page)
  • This page gave a $5,000-15,000 figure for the monthly cost of an additional attorney or paralegal. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says to use your own salary, benefits and overhead. (1 page)
  • This page gave a $5,000-15,000 figure for prior art search prices. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says there is no published price and to get a written quote. (1 page)
  • This page gave a $5,000-15,000 figure for an AI budget. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says the 1-3% of revenue is a planning assumption, not a measured benchmark, and the $5,000-15,000 is shown only as illustrative arithmetic on it. (1 page)
  • This page gave a $5,000-15,000 figure for monthly compliance monitoring retainers. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says retainers are the firm's own. (1 page)
  • This page gave a $5,000-15,000 figure for a typical mid-market deal's document count. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says document counts vary by deal. (1 page)
  • This page gave a $5,000-15,000 figure for a firm's annual cyber insurance premium and a 5-15% AI endorsement surcharge. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says there is no published figure and shows an illustration with its assumptions declared. (1 page)
  • This page gave a $5,000-15,000 figure for the cost of a paid vendor pilot (and a $150,000 'prevented mistake'). No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says to use the vendor's written pilot quote. (1 page)
  • This page gave a $5,000-15,000 figure for external AI audit assistance. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says to get written quotes. (1 page)
  • This page gave a $5,000-15,000 figure for yearly AWS infrastructure costs. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says costs depend on volume and can be estimated with the AWS Pricing Calculator. (1 page)
  • This page gave a $5,000-15,000 figure for a typical enterprise legal-AI procurement cycle. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says there is no published figure and to compare vendor quotes. (1 page)
  • This page gave a $5,000-15,000 figure for Filevine or SmartAdvocate implementation costs. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says to get a written quote from each vendor. (1 page)
  • This page gave a $5,000-15,000 figure for regulatory record size. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says page counts vary and should be counted per matter. (1 page)
  • This page gave a $5,000-15,000 figure for M&A data room size. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says size varies by deal. (1 page)
  • This page gave a $5,000-15,000 figure for the cost of implementing an AI program. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says to budget from your own quotes. (1 page)
  • This page gave a $5,000-15,000 figure for a typical sanction for a first offense with fabricated citations. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says amounts vary widely by court and conduct (the page keeps the dated, named amounts from its dataset). (1 page)
  • This page gave a $5,000-15,000 figure for an intake automation investment. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says to use your own quoted cost. (1 page)
  • This page gave a $5,000-15,000 figure for an AI-flagged document subset and its $10,000-25,000 cost. No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says there is no measured figure and to get quotes. (1 page)
  • This page gave a $5,000-15,000 figure for a jury consultant engagement (and a derived $10,000-25,000 saving). No vendor, regulator or survey we could find publishes it, so the figure was removed and the sentence now says to get a written quote from the consultant and the tool. (1 page)
  • This page said firms still staffing data rooms with 'armies of associates'. No source measures that, so the claim was removed and the sentence now says we have no measured figure and to pilot it on a closed deal. (1 page)
  • This page said firms still staffing due diligence with 'armies of junior associates' were overspending and under-covering. No source measures that, so the claim was removed and the sentence now says we have no measured figure and to pilot it on a closed deal. (1 page)
  • This page said compliance work 'required armies of associates reviewing policies manually'. No source measures that, so the claim was removed and the sentence now says we have no measured figure and to pilot it on a closed deal. (1 page)

Source

AIPLA, 2023 Report of the Economic Survey (page read November 2, 2026). It lists typical charges for law services in a members-only report and publishes no price or median for a prior art or patentability search, so the removed search-price range has no primary.; U.S. Bureau of Labor Statistics, Occupational Outlook Handbook, Paralegals and Legal Assistants (page read November 2, 2026). It gives median pay (May 2025: $62,890 a year), not a total monthly cost per hire, so the removed per-person range was replaced by a formula on the reader's own figures.

Pages that carry this correction (22)

Implementation timelines, contract terms and per-case hours with no vendor primary (Filevine '4-8 weeks' and '20+ hours per case', Docketwise '2-3 weeks', Imagility '4-8 weeks', Eve by Luminance '4-8 weeks' and 'annual commitments are standard') replaced by 'the vendor does not publish it' on 4 tool pages

· 4 pages

What changed

  • This page gave an implementation or onboarding timeline (4-8 weeks, 2-3 weeks, 2-3 months to full adoption), a contract term or a per-case hours saving (20+ hours) for a vendor that publishes none of them. The page now says the vendor does not publish it and tells readers to ask for it in writing; no range of ours replaces it. (4 pages)

Source

Filevine pricing page and FAQ, read November 2, 2026. All packages are custom and onboarding runs through Certified Implementation Partners; no implementation timeline, contract term or onboarding cost is published; Filevine MedChron page, read November 2, 2026. It says the tool saves hours of manual review and gives no per-case hours figure; Docketwise pricing page, read November 2, 2026. It lists tailored account setup and free data migration (annual plan) and no implementation timeline; Luminance home page, read November 2, 2026. It gives no implementation timeline or contract term. The Imagility site refused access on the same date, so no Imagility primary could be read

Pages that carry this correction (4)

Per-GB and monthly e-discovery prices for Everlaw, DISCO, Logikcull and Relativity replaced by 'no published rate' wording, and an unsourced 'under 15 attorneys' threshold labelled as our judgment, on 15 pages

· 15 pages

What changed

  • The AI document review for HR guide gave a per-GB monthly price range for Relativity. Relativity publishes no per-GB rate, so the figure was removed and the sentence says to get a written quote. (1 page)
  • The AI discovery review workflow page gave per-GB monthly prices for Everlaw and RelativityOne. Neither vendor publishes a per-GB rate, so both figures were removed and the page says to get written quotes. (1 page)
  • The best AI for document review page gave a per-GB starting price for Logikcull and said its pricing is flat per GB and that DISCO is competitive with Everlaw per GB. Logikcull lists a 10 GB minimum and no rate, and DISCO and Everlaw publish no rate, so these statements were replaced with what the pricing pages say. (1 page)
  • The best e-discovery AI tools page called DISCO's pricing competitive per GB and per user. DISCO publishes no rate, so the sentence now says to get a written quote. (1 page)
  • The Everlaw page gave a monthly base cost plus per-GB fees. Everlaw publishes no rate, so the figure and the per-GB model claim were removed and the page says it is priced by quote. (1 page)
  • The Logikcull page gave a monthly cost for Everlaw and a monthly range for Logikcull. Neither vendor publishes those figures, so they were removed; the page now says what Logikcull's pricing page lists (data stored, 10 GB minimum, no rate). (1 page)
  • The DISCO Cecilia page said DISCO has unlimited data processing and no per-GB surprises. DISCO's pricing page says it bills on processed data per GB, with unlimited Cecilia AI use at no additional charge, so the claim was replaced with that. (1 page)
  • The DISCO vs Relativity page gave monthly cost ranges for DISCO and for RelativityOne. Neither vendor publishes those figures, so they were removed and the pricing rows now say no rate is published. (1 page)
  • The Everlaw vs DISCO vs Logikcull page said Logikcull has no per-GB surprises and gave a monthly range for Everlaw. Logikcull bills on data stored with a 10 GB minimum and Everlaw publishes no rate, so both claims were replaced with what the pricing pages say. (1 page)
  • The DISCO pricing page said DISCO has all-inclusive pricing with no per-GB hosting charges. DISCO's pricing page says it bills on processed data per GB, so the wording was corrected, and the claim that Relativity charges extra for aiR was replaced with Relativity's own statement that AI capabilities come standard with committed data volume. (1 page)
  • The Relativity vs Everlaw 2026 page said DISCO has an all-inclusive model that eliminates per-GB surprises. DISCO bills per GB on processed data, so the sentence was corrected. (1 page)
  • The Relativity alternatives page said DISCO has all-inclusive pricing with no per-GB surprises. DISCO bills per GB on processed data and publishes no rate, so the sentence was replaced with what its pricing page says. (1 page)
  • The switching from Relativity to Everlaw guide gave monthly and annual cost ranges for both platforms, a saving and a migration cost per matter, and said Everlaw publishes its pricing model. Neither vendor publishes a rate, so the figures were removed and the page says to compare written quotes. (1 page)
  • The Logikcull vs Relativity guide described flat-rate Logikcull pricing and, in its source, per-GB prices for both vendors. Logikcull bills on data stored (10 GB minimum) and neither vendor publishes a rate, so the pricing statements now say that and recommend written quotes. (1 page)
  • The eImmigration page said the all-in-one trade-off is right for most firms under 15 attorneys. We found no source for that threshold, so the page now says under about 15 attorneys and labels it our judgment, not a measured threshold. (1 page)

Source

Everlaw, pricing page (everlaw.com/pricing, read November 2, 2026): lists features and a demo request, no price or per-GB rate; DISCO, pricing page (csdisco.com/pricing, read November 2, 2026): says it bills on processed data per GB with unlimited Cecilia AI use at no additional charge, and offers a quote; no amount; Logikcull, pricing page (logikcull.com/pricing, read November 2, 2026): pay as you go on data stored, 10 GB minimum, unlimited users; no dollar amount; Relativity, pricing page (relativity.com/platform/pricing, read November 2, 2026): pay as you go or a one- or three-year commitment on committed data volume, AI capabilities standard; request pricing, no amount

Pages that carry this correction (15)

Claude Mythos 'no major AI company has done before' and the Glasswing partner description replaced with Anthropic's own wording

· 1 page

What changed

  • The Claude Mythos page said that Anthropic did something no major AI company has done before and described Project Glasswing as involving 12 of the largest technology and financial companies. Anthropic's announcement does not claim a first, and its twelve launch partners include Anthropic itself and the Linux Foundation, so the page now reports what Anthropic said (it does not plan to make Mythos Preview generally available) and names launch partners from the announcement. (1 page)

Source

Anthropic, Project Glasswing announcement (April 7, 2026, read November 2, 2026): 'We do not plan to make Claude Mythos Preview generally available'; lists twelve launch partners (including Anthropic and the Linux Foundation) and over 40 additional organizations

Pages that carry this correction (1)

An unsourced 'months of manual tabulation now takes days' removed, and 'courts and regulators require redaction documentation' replaced by what Federal Rule of Civil Procedure 5.2 actually requires

· 2 pages

What changed

  • The government contractor legal page said a task that used to take a claims consultant months of manual tabulation now takes days. We found no published figure for either duration, so the page now says we have no measured figure and suggests testing the tool on one closed claim. (1 page)
  • The document redaction page said courts and regulators increasingly require documentation of your redaction methodology. Rule 5.2 of the Federal Rules of Civil Procedure requires redaction of listed identifiers in filings but does not require a log, so the page now cites the rule and presents the log as your own record. (1 page)

Source

Federal Rule of Civil Procedure 5.2, Privacy Protection for Filings Made with the Court (Cornell Legal Information Institute, read November 2, 2026), requires redaction of listed identifiers and allows an optional sealed reference list, with no log requirement

Pages that carry this correction (2)

Ten pages that named superseded models as the current ones behind a plan or a tool, and the verdicts that rested on a test of those versions

· 10 pages

What changed

  • Replaced Claude 3.5 Sonnet and Opus 4 as the models of the Claude plans on three comparison pages. Anthropic's pricing page lists Sonnet and Haiku on the free tier and Opus, Sonnet and Haiku on Pro and Team; we also removed the 'best accuracy-to-cost ratio in the market' line, which had no measurement. (3 pages)
  • Removed GPT-4o and GPT-4o mini as the models of ChatGPT Free and Plus on two comparison pages. OpenAI's plan pages answered 403 when we tried to read them, so the pages now say we have not verified the model or limits. (2 pages)
  • Removed GPT-4 as the model behind CoCounsel and Harvey. Thomson Reuters' CoCounsel page and Harvey's security page name no underlying model, so the pages say that instead; two passages that pointed to GPT-4 as the cheap alternative now say a general model such as Claude. (5 pages)
  • On the Claude vs Gemini legal writing page, removed the result and verdicts that named Claude 3.7 Sonnet, Gemini 1.5 Pro and Opus 4: we have no record of the model versions or scoring behind them. The page now says there is no current test, and lists the current windows from each provider's docs. (1 page)
  • Replaced the 200K-token window given for Claude on three pages. Anthropic's model docs list a 1M-token window for its current models; the pages say we have not checked the limit in the app. (3 pages)
  • Removed 'best writing quality' and 'writes better legal prose' rankings of Claude on two pages. No source or current test supports them; the pages ask the reader to run both tools on their own drafts. (2 pages)

Source

Anthropic, models overview (platform.claude.com, read November 2, 2026): current models are Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, each with a 1M-token context window; Opus 4.x and Sonnet 4.6 are listed as legacy; Anthropic, pricing (claude.com/pricing, read November 2, 2026): Free lists Sonnet and Haiku; Pro and Team list Opus, Sonnet and Haiku; Google, Gemini API models (ai.google.dev, read November 2, 2026): Gemini 4 Argon with a 1M context window; the 2.5 models limited to users who have used them before; Harvey, security (harvey.ai/security, read November 2, 2026): refers to external model providers without naming a model; Thomson Reuters, CoCounsel Legal (read November 2, 2026): names no underlying model

Pages that carry this correction (10)

GPT-4, GPT-4o, o1 and 'Codex 5.4' named as what ChatGPT, CoCounsel or Harvey run, and the Gemini 3.1 Pro context-window verdict, replaced on 17 pages

· 17 pages

What changed

  • Removed GPT-4 and GPT-4o as the model behind ChatGPT in tool lists, examples and verdicts, and the claims that GPT-4o hallucinates citations at a given rate or that ChatGPT underperforms Claude on legal reasoning. OpenAI's models page lists GPT-6 models as current and does not list GPT-4 or GPT-4o, and we have no current comparison, so the pages now say ChatGPT or OpenAI models, or that we have not verified the comparison. (12 pages)
  • AI legal glossary: replaced GPT-4 as the example model in the definitions, removed its 128K-token context window and its 'broadest ecosystem' and 'higher hallucination rate' labels (no current source), and now say we have not verified OpenAI's current context window. (1 page)
  • Removed GPT-4-class models as the model behind CoCounsel and Harvey's compliance module. Thomson Reuters's CoCounsel Legal page names no underlying model and Harvey's security page names none, so the pages say that. (2 pages)
  • Removed 'GPT-5.5 powers Codex 5.4': OpenAI has no model page named Codex 5.4 (gpt-5.4-codex returned 404), so the page says we have not verified which OpenAI model Codex runs. On the Gemini 3.1 Pro comparison, replaced 'multi-million-token context window' with the 1,048,576-token input limit Google lists, and removed the claim that Gemini's window beats Opus 4.7's, which no source supported. Google's pricing page lists Gemini 3.1 Pro Preview, which the page's status block already cited. (2 pages)

Source

OpenAI, API models (developers.openai.com, read November 2, 2026): current models GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna; GPT-4o and GPT-4 are not listed; the model page for gpt-5.4-codex returned 404; Google, Gemini API models (ai.google.dev, read November 2, 2026): flagship Gemini 4 Argon with a 1M-token context window; gemini-3.1-pro-preview listed; Google, Gemini API pricing (ai.google.dev, read November 2, 2026): Gemini 3.1 Pro Preview is priced per 1M tokens, with a higher tier for prompts above 200k tokens; Anthropic, models overview (platform.claude.com, read November 2, 2026): current Claude models list a 1M-token context window; Thomson Reuters, CoCounsel Legal (legal.thomsonreuters.com, read November 2, 2026): names no underlying model

Pages that carry this correction (17)

GPT-4, GPT-4o, o1 and Gemini 1.5 named as what ChatGPT Enterprise, Copilot, Harvey, Spellbook or Gemini run, and verdicts built on them, replaced on 15 pages

· 15 pages

What changed

  • Removed GPT-4o, o1 and GPT-4 as the model behind ChatGPT's Plus and Enterprise plans. OpenAI's models page lists GPT-6 models as current and does not list GPT-4o or GPT-4, and the plan pages answered 403 earlier today, so the pages now say OpenAI models or that we have not verified which model a plan runs. (6 pages)
  • Removed GPT-4 or 'GPT-based' as the model Legalese Decoder, Harvey and Spellbook run today, and the reported Harvey switch to Claude. Harvey's security page and Spellbook's site name no model, so the pages state that and say we have not verified it. (3 pages)
  • Replaced Gemini 1.5 Pro (and Gemini 2.5 Pro, now deprecated) with Google's current flagship, Gemini 4 Argon, and its 1M-token context window as Google lists it. The page had said Claude's context ceiling was lower than Gemini's; Anthropic's model docs list 1M tokens for Claude as well, so that claim was removed. The Gemini 2.5 Pro legal-capability and cost verdicts were removed (no primary). (2 pages)
  • Removed verdicts that rested on GPT-4 or GPT-4o as the comparison model and had no traceable source: benchmark gaps of 3-7% and 8-12%, a 67% attorney preference, a 'low 80s' BigLaw Bench score, a cost-per-token edge, language and multimodal advantages, and a higher hallucination rate. GPT-4 and GPT-4o are no longer in OpenAI's current model list; the pages now say we have not verified a current comparison. (4 pages)
  • Replaced GPT-4o as ChatGPT's context limit, as the OpenAI API's model, as the model behind Microsoft Copilot, as the tool to build coverage-opinion workflows on (insurance-defense page, now 'Claude or the OpenAI API') and as the ChatGPT entry in a list of vendors whose data terms changed with OpenAI's models in general or with ChatGPT (Microsoft's Copilot page lists the latest OpenAI models as GPT-5.X), or with 'not verified'. The consumer-plan price attached to ChatGPT on the Spellbook vs Harvey page was dropped with the model name. (6 pages)

Source

OpenAI, API models (developers.openai.com, read November 2, 2026): current models GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna; GPT-4o and GPT-4 are not listed; Google, Gemini API models (ai.google.dev, read November 2, 2026): flagship Gemini 4 Argon with a 1M-token context window; gemini-2.5-pro is marked deprecated; Anthropic, models overview (platform.claude.com, read November 2, 2026): current Claude models list a 1M-token context window; Microsoft, Copilot pricing (microsoft.com, read November 2, 2026): lists the latest OpenAI models as GPT-5.X

Pages that carry this correction (15)

GPT-4o, GPT-4 and o1 named as what a ChatGPT plan, Spellbook, CoCounsel or Westlaw Precision runs, replaced on 15 pages

· 15 pages

What changed

  • Removed GPT-4o and o1 as the models behind ChatGPT's plans (and GPT-4o behind Perplexity Pro). OpenAI's models page lists GPT-6 models as current and does not list GPT-4o; OpenAI's and Perplexity's plan pages answered 403 and were not worked around, so the pages now say we have not verified which model each plan runs. (6 pages)
  • Removed GPT-4 (and GPT-4 plus Claude) as the model Spellbook, CoCounsel, Harvey and Westlaw Precision are built on. Spellbook's and Thomson Reuters' pages, and Harvey's security page, name no model, so the pages no longer name one. (5 pages)
  • Replaced GPT-4 and GPT-4o as the model to build on or recommend with a general reference to Claude or OpenAI's models (on the boutique-firms page, with 'a general-purpose assistant such as Claude' and without the monthly price given for ChatGPT, which we could not verify: OpenAI's plan pages answered 403). GPT-4 and GPT-4o are no longer in OpenAI's current model list. (5 pages)
  • Removed the statement that ChatGPT's training data runs through early 2024 (it described GPT-4). We have not verified the current model's cutoff. (1 page)

Source

OpenAI, API models (developers.openai.com, read November 2, 2026): current models GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna; GPT-4o and GPT-4 are not listed; Thomson Reuters, CoCounsel Legal (read November 2, 2026): names no underlying model; Spellbook (spellbook.com, read November 2, 2026): refers only to its 'LLM providers', names no model

Pages that carry this correction (15)

Pages dedicated to one model now carry a dated status block, and the best-model verdicts and wrong GPT-5.4 figures they held were rewritten

· 6 pages

What changed

  • The 87.6% SWE-bench Verified score is now described as the figure Anthropic reported at release, not as the highest of any model, and the claim that Opus 4.7 is the better coding agent became a candidate to trial: newer models have shipped and we have not compared it with Codex. (3 pages)
  • Removed 'both are flagship models' from the Opus 4.7 vs GPT-5.5 comparison: Anthropic lists Opus 4.7 as a legacy model, and OpenAI's page for GPT-5.5 shows no legacy notice while its model index leads with GPT-6 Astra. (1 page)
  • Removed the verdicts that Claude Opus 4.7 writes better legal documents than GPT-5.4 or Gemini 3.1 Pro, at lower cost or in fewer revisions. We have no current test of either, so the pages now say that and suggest running both on your own documents. (2 pages)
  • Corrected the GPT-5.4 figures on the Opus 4.7 vs ChatGPT page: OpenAI lists $2.50 per million input tokens and $15 output (not $7.50 and $30) and a 1,050,000-token context window (128,000 is the maximum output, not the window). The $1,200 a year seat saving was removed: Claude Team and ChatGPT Business list the same seat prices. (1 page)
  • The 90.9% and 84.2% BigLaw Bench scores in this page's headline are not published on any vendor page we read, so the body now calls them unverified instead of presenting them as the gap between a draft that needs two redlines and one that does not. (1 page)

Source

Anthropic, Claude Opus 4.7 (platform.claude.com, read November 2, 2026): legacy, released April 16, 2026, $5 / $25 per million tokens; Opus 5.5 is the current Opus model; Anthropic, Claude Sonnet 4.6 and Sonnet 5.5 (platform.claude.com, read November 2, 2026): Sonnet 4.6 legacy, Sonnet 5.5 "Active (latest)", released September 28, 2026; OpenAI, GPT-5.4 model page (developers.openai.com, read November 2, 2026): $2.50 / $15 per million tokens, 1,050,000-token context window, 128,000 maximum output; OpenAI, GPT-5.5 model page and model index (read November 2, 2026): GPT-5.5 shows no legacy or deprecation notice; the index leads with GPT-6 Astra; Google, Gemini API deprecations (aistudio.google.com, read November 2, 2026): gemini-3.1-pro-preview released February 19, 2026, no shutdown date announced

Pages that carry this correction (6)

Unsourced 'in minutes instead of hours' claims and before/after task times ('takes 4-6 hours, takes 30 minutes with AI') removed or rewritten as qualitative wording on 51 practice, tool, comparison, guide, industry and use-case pages

· 51 pages

What changed

  • This page said the work took 'minutes instead of hours' or went 'from hours to minutes'. No source supported a measured time, so the sentence now says the work is faster than doing it by hand, or tells readers to time their own. (33 pages)
  • This page gave before-and-after task times for AI-assisted work (for example 'takes 4-6 hours, takes 30 minutes with AI'). No source supported them, so they were removed and the page tells readers to time their own work. (18 pages)
  • This page gave a before-and-after task time or staffing figure for AI-assisted work. No source supported it, so it was removed and the page tells readers to time their own work. (5 pages)

Source

Thomson Reuters Institute, Making the most of AI's potential time savings (September 12, 2024; page read November 2, 2026). It reports lawyers' own estimate of 4 hours a week saved from AI in general, a survey estimate and not a measurement; none of the removed task figures has a primary, so none was replaced by another number

Pages that carry this correction (51)

Case valuation page: Lex Machina '70-80% accuracy' removed and Darrow '5 million+ potential violations' corrected to the vendor's own '5M+ signals detected monthly'

· 1 page

What changed

  • Removed the statement that Lex Machina's predictions achieve '70-80% accuracy': neither Lex Machina's product page nor any other primary we read publishes an accuracy figure. Darrow's '5 million+ potential violations per month' now reads as Darrow states it, '5M+ signals detected monthly' (signals, not confirmed violations), with a link and the read date. The '3,000+ US attorneys on the platform' claim, which Darrow does not state, was replaced by Darrow's own '22K+ legal professionals'. The FAQ about Lex Machina accuracy was rewritten to say no figure is published. (1 page)

Source

Lex Machina product page on LexisNexis (read November 2, 2026, analytics on judges, lawyers, parties and courts; no accuracy percentage for outcome predictions; Darrow homepage (read November 2, 2026, '5M+ signals detected monthly' and '22K+ legal professionals powered by Darrow's intelligence'; no '3,000+ US attorneys' and no count of violations

Pages that carry this correction (1)

Lexis+ AI 'cannot hallucinate' replaced with the vendor claim and the Stanford measurement, Claude Pro priced with both billing terms, Doctrine pricing page no longer described as blocked

· 5 pages

What changed

  • Two Lexis+ AI comparison pages (vs Claude, vs Harvey AI) said Lexis+ AI cannot hallucinate case citations, will not hallucinate case names, has guaranteed citation accuracy, or that its database grounding eliminates the hallucination problem. LexisNexis markets its citations as hallucination-free, but a preregistered Stanford study (arxiv 2405.20362) found its research tools and Thomson Reuters's each hallucinated between 17% and 33% of the time, less than a general chatbot but not none. The pages now attribute the claim to LexisNexis, report what Stanford measured and tell readers to check every citation. (2 pages)
  • Three pages (Lexis+ AI vs Claude, Spellbook vs Claude, Gemini vs Claude) gave Claude Pro as a flat $20 a month. Anthropic's pricing page lists $17 a month with the annual discount and $20 billed monthly. The pages now give both prices with the date they were read. (3 pages)
  • The Doctrine guide said Doctrine's pricing page was blocked and gave euro price ranges attributed to contract leaks and press. The page loads (HTTP 200, read November 2, 2026), is in French, offers a 7-day free trial with no card and shows no price in its text. The guide now says that and removes the unsourced ranges. (1 page)

Source

Magesh et al., 'Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools' (Stanford RegLab and HAI, arxiv 2405.20362, abstract read November 2, 2026): the LexisNexis (Lexis+ AI) and Thomson Reuters tools each hallucinate between 17% and 33% of the time, less than GPT-4 but not none; the paper quotes LexisNexis (2023) as guaranteeing 'hallucination-free' legal citations; Anthropic pricing (claude.com/pricing, read November 2, 2026): Pro $17 per month with the annual subscription discount, $20 if billed monthly; Doctrine pricing page (doctrine.fr/tarifs, read November 2, 2026): loads (HTTP 200) in French with a 7-day free trial, no card, prices listed excluding tax and no price in its text

Pages that carry this correction (5)

Legalfly homepage headline quoted correctly; the unsourced 'over 35 state bars' vLex count confirmed as already replaced

· 1 page

What changed

  • The Legalfly review quoted the company's homepage as calling it 'legal AI built for the whole enterprise'. Legalfly's homepage headline reads 'legal AI for the whole business', so the quote now uses that wording and the date we read it. (1 page)

Source

Legalfly homepage (read November 2, 2026), whose headline reads 'legal AI for the whole business', not 'legal AI built for the whole enterprise'; vLex, Fastcase Library (read November 2, 2026), which says '80+ state, county, and specialty bar associations' and gives no count of bars offering Vincent AI

Pages that carry this correction (1)

Kansas court-rules page no longer says the state has no AI committee; Kansas Supreme Court Order 2025-CM-017 and its Ad Hoc AI Committee added

· 1 page

What changed

  • This page said Kansas had no task force, committee or administrative order on AI and no movement toward attorney AI rules. The Kansas Supreme Court created an Ad Hoc Artificial Intelligence Committee by Administrative Order 2025-CM-017 in February 2025, which secondary reports say is charged with recommending policies on attorney AI use. The page now says what we found (no Kansas disclosure rule or bar ethics opinion that we could confirm) and points to the order. (1 page)

Source

Kansas Supreme Court, Administrative Order 2025-CM-017, Creation of Ad Hoc Artificial Intelligence Committee (dated Feb. 27, 2025; read through secondary reports, confirm at the order)

Pages that carry this correction (1)

Juro page: three unsourced claims removed and the AI feature renamed to the name Juro uses

· 1 page

What changed

  • Removed the statement that Juro solves the problem for '90% of companies' that enterprise tools ignore. The page gave no source and we found no published count of how many companies that is. It now cites Juro's own pricing-page wording about business users self-serving on contracts and gives no figure. (1 page)
  • Removed the claim that Juro delivers about four-fifths of an enterprise CLM's value for a tenth of its cost. Juro publishes no price, so no such ratio can be computed. The FAQ now says we cannot state it and suggests testing on your own templates before comparing quotes. (1 page)
  • Removed the claim that Juro's AI handles 80% of routine review and that teams use it to shorten first review. Juro publishes no such share. The answer now says what juro.com/ai describes (review of third-party contracts against your playbooks, redlines, escalation to legal when something falls outside the rules) and that we have not tested it. (1 page)
  • The page called Juro's AI feature 'Review Agent' and said it 'brings genuine AI value at a price point that doesn't require board approval'. Juro's own pages call it AI Review (pricing page) and describe it as a contract review agent (juro.com/ai); no price is published. The page now uses AI Review, attributes the description to Juro with the read date, and drops the price-point claim. The integrations answer was also corrected to the list on Juro's pricing page (Slack, Google Drive, Companies House; Salesforce, HubSpot and Workday at extra cost). (1 page)

Source

Juro, pricing page (read November 2, 2026), lists AI Review as 'Auto-redline against custom playbooks', publishes no price and says all plans include unlimited users; Juro, AI page (read November 2, 2026), describes its contract review agent, which reviews third-party contracts against your playbooks and redlines them in Juro, Slack, Microsoft Teams and Word; Juro, home page (read November 2, 2026), headline 'Review and redline contracts with AI agents'

Pages that carry this correction (1)

Judge pages (Lin, Ellis, Noreika, Torres, Mizelle, Reeves, Preska, Salas, Summerhays, Corley): wording re-read against the courts' own orders and the Federal Judicial Center

· 10 pages

What changed

  • The Lin page said she is presiding over xAI v. OpenAI, calling it a central corporate battle that makes her courtroom uniquely AI-aware. Her Civil Standing Order (September 4, 2026) is unchanged and was re-read; the court's cases-of-interest list does not include the case, and press reports describe a trade-secrets suit (No. 3:25-cv-08133) that she dismissed in June 2026. The page now says so, attributes that to press reports and drops the unsourced characterizations. (1 page)
  • The Ellis page said several N.D. Ill. federal judges require AI disclosure and told readers to follow the strictest approach. None of the court's judge pages requires disclosure: Coleman, Johnston and Valdez point to Rule 11 and sanctions, Magistrate Judge Fuentes's March 3, 2026 order refers to the Illinois Supreme Court policy, and Judge Ellis's own page has no AI provision. The page now says so and labels the disclosure advice as ours. (1 page)
  • The Noreika page described her docket as corporate and bankruptcy heavy, called Hatch-Waxman a Delaware specialty with stakes in the hundreds of millions, and predicted AI disclosure rules from Delaware's transparency culture. Her chambers page lists patent scheduling orders and no AI provision; Chief Judge Connolly's disclosure orders apply to cases assigned to him and say nothing about AI. Those statements are removed or labelled as advice. (1 page)
  • The Torres page matched her Individual Practices (revised November 5, 2025, section VII). Two characterizations without a primary source (the Ripple ruling "moved crypto markets" and a docket she "sees daily") are removed. (1 page)
  • The Mizelle page said she was the youngest person Trump appointed, that airlines dropped mask rules within hours of her ruling, gave ABA-rating details and said filers should prepare as if the Supreme Court were reading over their shoulder. The court's pages and the Federal Judicial Center show her commission on November 20, 2020 and the ABA rating only, and the district's list shows AI orders from other judges but none from her. Press-only items are now attributed to press reports; the rest is removed. (1 page)
  • The Reeves page called his district ground zero for the judicial AI accountability crisis, said formal requirements are coming and cited his schooling and rulings without sources. His page, the local rules and the administrative procedures contain no AI provision; the Wingate matter rests on press reports and Senator Grassley's disclosures as reported, with no Senate or court primary cited. The page now says so and keeps only what the Federal Judicial Center confirms. (1 page)
  • The Preska page tied her to the Maxwell prosecution and described her instincts and values. Her Individual Practices (April 1, 2025) and the Federal Judicial Center (senior status March 1, 2017) support neither; they contain no AI requirement. Those statements are removed and the page now describes what the practices set (page limits, letters, sealing, sentencing). (1 page)
  • The Salas page dated New Jersey's Daniel's Law to 2021 and described the federal law imprecisely, and also gave unsourced honors and firsts. The state law was approved November 20, 2020 (P.L. 2020, c.125); her procedures and the district's standing orders contain no AI provision. Corrected, with the unsourced items removed or attributed to press reports. (1 page)
  • The Summerhays page omitted his years as a bankruptcy judge, described the Title 42 order as a restraining order, and cited a conference talk and personal traits without sources. The Federal Judicial Center lists him as W.D. La. bankruptcy judge 2006 to 2018; press reports describe a preliminary injunction of May 20, 2022. His chambers page and the district's local rules contain no AI provision. Corrected or removed. (1 page)
  • The Corley page gave 2021 as the year she became a district judge and claimed knowledge and docket traits without sources. The Federal Judicial Center confirms March 2022 (magistrate judge from 2011); her civil and criminal standing orders contain no AI provision, while nine other N.D. Cal. judges' orders do. Corrected and the unsourced statements removed. (1 page)

Source

N.D. Cal., Judge Lin's Civil Standing Order dated September 4, 2026 (read November 2, 2026): generative AI paragraph, no disclosure requirement; N.D. Ill., judge pages for Ellis, Coleman, Johnston and Valdez and Magistrate Judge Fuentes's civil standing order dated March 3, 2026 (read November 2, 2026); D. Del., Judge Noreika's chambers page and posted orders (read November 2, 2026): no AI provision; S.D.N.Y., Judge Torres's Individual Practices in Civil Cases, revised November 5, 2025, section VII (read November 2, 2026); M.D. Fla., Judge Mizelle's judge page and the court's standing-orders page (read November 2, 2026): no AI order from her; orders from other judges listed; S.D. Miss., Judge Reeves's page, local rules and administrative procedures (read November 2, 2026): no AI provision; S.D.N.Y., Judge Preska's Individual Practices dated April 1, 2025 (read November 2, 2026): no AI provision; D.N.J., Judge Salas's procedures and the district's standing orders (read November 2, 2026): no AI provision; W.D. La., Judge Summerhays's chambers page and the district's local rules effective May 7, 2024 (read November 2, 2026): no AI provision; N.D. Cal., Judge Corley's Civil Standing Order (revised January 21, 2025) and Criminal Standing Order (March 2023), read November 2, 2026; Federal Judicial Center, biographical directory of federal judges (judges.csv, read November 2, 2026)

Pages that carry this correction (10)

Judge pages (Wilken, Kollar-Kotelly, Suddaby, Rakoff, Koeltl, Kang, Chhabria, Lauck, Lamberth, Wang): wording re-read against the courts' own orders and the Federal Judicial Center

· 9 pages

What changed

  • The Wilken page called the Northern District of California the epicenter of AI litigation and O'Bannon a groundbreaking win; unsourced superlatives were removed, O'Bannon is now described as affirmed in part and reversed in part by the Ninth Circuit, and the Noland sanction is credited to the Court of Appeal, not the State Bar. (1 page)
  • The Kollar-Kotelly page gave her FISA Court service as 2002 to 2013 and dated a surveillance opinion to August 2004; the Federal Judicial Center lists 2002 to 2009 and press dates the opinion July 14, 2004. Superlatives were removed and the DOGE order is described as interim. (1 page)
  • The Suddaby page described him as an active district judge; the Federal Judicial Center records senior status on September 1, 2024. An unsourced sentence about other New York judges was removed. (1 page)
  • The Rakoff page opened with an unsourced superlative; it now states that his June 3, 2024 Individual Rules contain no AI text and that he took senior status on December 31, 2010. (1 page)
  • The Koeltl page called him a senior judge, described his docket and the S.D.N.Y. as 'ground zero' without a source; the Federal Judicial Center lists no senior status, his March 17, 2025 Individual Practices have no AI provision, and the Mata sanction is stated as in the order (US$5,000, jointly and severally on the two attorneys and their firm). (1 page)
  • The Chhabria page called the Meta case Silverman v. Meta and described it as a class action; the docket caption is Kadrey v. Meta Platforms, No. 3:23-cv-03417-VC. His October 26, 2026 civil standing order has no AI filing requirement; unsourced claims about his expertise were removed. (1 page)
  • The Lauck page now quotes the September 1, 2026 order's AI paragraph and gives her chief judge date (December 4, 2025); unsourced pace, caseload and agency claims were removed. (1 page)
  • The Lamberth page attributed an AI task force to the D.D.C. (it belongs to the D.C. Courts), called his tenure 37 years and listed unsourced claims about his docket; his dates now follow the Federal Judicial Center and the Grassley inquiry is limited to what the Senate Judiciary release of October 6, 2025 says. (1 page)
  • The Nina Wang page paraphrased the order's certification as 'personal human review' and carried an internal drafting note ('Coder handoff') in its body; the wording now follows the order, the table names her as a U.S. District Judge (D. Colo.), and the note was removed. (1 page)

Source

Federal Judicial Center, Biographical Directory of Article III Federal Judges (judges.csv, read November 2, 2026): appointment, senior status and service dates; N.D. Cal., Judge Chhabria's Civil Standing Order (dated October 26, 2026) and Judge Kang's Civil Standing Order (July 16, 2025), read November 2, 2026; S.D.N.Y., Individual Rules of Judge Rakoff (June 3, 2024) and Individual Practices of Judge Koeltl (March 17, 2025), read November 2, 2026; E.D. Va., Standing Orders (Chief Judge Lauck, September 1, 2026), read November 2, 2026; N.D.N.Y. Local Rules 2026 and General Orders, read November 2, 2026

Pages that carry this correction (9)

Judge pages (Beetlestone, Engelmayer, Alsup, Cote, Nathan, Castel, Berman Jackson, Boasberg, Andrews): wording re-read against the courts' own orders and the Federal Judicial Center

· 9 pages

What changed

  • The Alsup page described him as an active judge with AI disclosure requirements, called him the most technically sophisticated federal judge and said he taught himself Java. The court's December 15, 2025 notice says he took inactive status before the end of December 2025 (the Federal Judicial Center gives senior status on January 21, 2021); we found no AI order of his, and the court's Standing Order for All Judges (November 30, 2023) does not mention AI. The disclosure requirements, superlatives and the hours-of-reading claim had no source and are removed. The Meta case is Kadrey v. Meta Platforms (No. 3:23-cv-03417-VC), not Silverman v. Meta. (1 page)
  • The Beetlestone page said two E.D. Pa. judges have AI standing orders. A third, Judge Marston, posts a generative AI order on her chambers page that requires a certification naming the tool and how it was used. The page also said the Pennsylvania Bar Association issued Formal Opinion 2024-200; it is a joint opinion of the Pennsylvania and Philadelphia Bar Associations and is advisory. The claim that the court is one of the busiest in the country had no source and is removed. Judge Beetlestone's own Policies and Procedures do not mention AI. (1 page)
  • The Engelmayer page said the S.D.N.Y. produces a large share of AI case law, named his former law firm and predicted that a fabricated authority would be noticed. The Federal Judicial Center lists his prosecutor and Solicitor General service and private practice from 2000 to 2011 without a firm; the first and last claims had no source. His Rule 3.H (May 2026) was re-read and is unchanged. The Heppner privilege ruling is now described as reported in law-firm summaries because the docket could not be opened. (1 page)
  • The Cote page called her one of the most experienced judges on the bench and cited construction litigation over the Baruch College campus. The Federal Judicial Center records senior status on December 15, 2011, and we found no source for the Baruch matter or the characterizations, so they are removed. A further opinion in AFGE v. OPM (October 20, 2026) is now noted. Her civil and criminal Individual Practices contain no AI provision. (1 page)
  • The Nathan page called her best known for the Maxwell trial and Park v. Kim the Second Circuit's best-known AI decision, and said the court does not work through individual judges' orders. None of this had a source we could read, so the Maxwell role is attributed to press and the other statements are removed. Her appointment dates were checked against the Federal Judicial Center and are unchanged. (1 page)
  • The Castel page said the Mata v. Avianca sanctions decision is why many judges wrote AI rules. We found no primary source for that causation, so it is attributed to press and marked unverified. His senior status (August 5, 2017) is now stated. His civil Individual Practices (amended March 17, 2025) contain no AI provision. (1 page)
  • The Berman Jackson page gave her senior status as 2024; the Federal Judicial Center says May 1, 2023. It also called her one of the most high-profile judges, described the quality of opposing counsel and Department of Justice policy, and predicted referral to the DC Bar over an AI error. None of that had a primary source and it is removed. We found no AI order from her. (1 page)
  • The Boasberg page quoted him as saying lawyers must disclose AI use and said he sanctioned an attorney for undisclosed AI use. IAPP's report of the 2026 panel and dig.watch attribute that quotation to Judge Allison Burroughs. IAPP reports that Boasberg required a party to pay the other side's fees after AI-generated errors in briefs, not for undisclosed use, and cites no order. The DC Courts AI Task Force it credited to him belongs to the local D.C. Courts, not the federal district court. We found no AI order from him. (1 page)
  • The Andrews page presented him as an active judge; he took senior status on January 1, 2024. Statements on Delaware's caseload and on where hallucination is common had no source and are removed or softened. His two posted standing orders (briefing in all cases and defense motions) were re-read and do not mention AI. (1 page)

Source

Federal Judicial Center, Biographical Directory of Article III Federal Judges (judges.csv, read November 2, 2026): appointment, senior status and resignation dates; N.D. Cal., public notice 'Judge William Alsup Status Change' (December 15, 2025), read November 2, 2026; E.D. Pa., Judge Marston's generative AI standing order, and Judge Beetlestone's Policies and Procedures, read November 2, 2026; S.D.N.Y., Judge Engelmayer's Individual Rules (May 2026), Rule 3.H, and Judge Cote's Individual Practices in Civil and Criminal Cases, read November 2, 2026; IAPP, report on the 2026 Global Summit panel of federal judges (attributes the disclosure quotation to Judge Burroughs), read November 2, 2026; D.D.C., Judge Boasberg's biography and the court's standing orders page, read November 2, 2026; D. Del., Judge Andrews' court page and posted standing orders, read November 2, 2026

Pages that carry this correction (9)

Judge pages (Broderick, Kelly, Caproni, Kennelly, Chen, Doughty, Birotte, Howell, Walker): wording re-read against the courts' own orders and the Federal Judicial Center

· 9 pages

What changed

  • The Broderick page said he was Deputy Commissioner for Legal Matters at the NYPD; the Federal Judicial Center lists no such post (that title belonged to a different Broderick decades earlier). His career per the Center: private practice, Assistant U.S. Attorney in the S.D.N.Y. (1994 to 2002, chief of the Violent Gangs Unit from 1999) and private practice again until 2013. 'One of the clearest AI rules in the district' and the unsourced caseload sentence are removed; the Section 4.J disclosure and certification rule is unchanged (Civil Rules revised October 29, 2025, re-read November 2, 2026). (1 page)
  • The Kelly page presented the 'DC Courts AI Task Force' as the D.D.C.'s own framework and predicted formal D.D.C. attorney rules from it. The Task Force belongs to the local D.C. courts, separate from the U.S. District Court, whose Local Rules (September 2026) and Judge Kelly's standing order template (revised March 2021) contain no AI provision. Also removed: a Baker Botts line not in his biography or the Federal Judicial Center record, the claim that the S.D.N.Y. has a district-wide AI certification rule, and unsourced docket and temperament statements. (1 page)
  • The Caproni page omitted that she took senior status on January 7, 2025 (Federal Judicial Center) and described her Individual Practices as emphasizing accuracy of citations; her civil and criminal practices (revised March 17, 2025) and pro se practices contain no AI provision. Unsourced statements about her FISA and national-security docket, the Mata v. Avianca effect on the bench and 'highest briefing standards' are removed. (1 page)
  • The Kennelly page said multiple N.D. Ill. judges require AI disclosure and advised treating disclosure as expected practice. The Fuentes, Johnston and Coleman orders point to Rule 11 and verification and do not require disclosure; Judge Cole's order was not available. Unsourced statements that he sits by designation in Delaware, speaks at mass-tort conferences, manages cases efficiently, or ruled in specific DEI and ICE cases are removed or attributed to press coverage. (1 page)
  • The Chen page said she is the first openly LGBTQ+ Asian-American federal judge in U.S. history, listed Grimm, FIFA and other cases she presided over, and described anti-trafficking work; none of this is in the court's biography or the Federal Judicial Center record, so it is removed or attributed to press coverage. Her Individual Practices (updated October 2026) contain no AI provision, and her DOJ service is now dated as the Center lists it (1991 to 2007, 2008 to 2013). (1 page)
  • The Doughty page said the Supreme Court stayed portions of the Missouri v. Biden injunction, gave a 162-page length, a July 4, 2023 date and a quotation about 'the most massive attack on free speech' without a source it had read. Murthy v. Missouri (June 26, 2024) says the Court stayed the injunction, then held the plaintiffs lacked standing; the unread details are removed. His career per the Federal Judicial Center: private practice in Rayville, Louisiana, a state district judgeship (2009 to 2018), chief judge since 2022. (1 page)
  • The Birotte page called the Central District 'the largest district in the country', gave sanctions amounts from other cases without a source it had read and said the PDF of his civil standing order could not be reopened. We re-read his Standing Order in Civil Cases (March 15, 2024), Scheduling Conference order and Civil Jury/Court Trial order (September 4, 2026) on the court's site on November 2, 2026; none mentions generative AI. (1 page)
  • The Howell page said she was chief counsel to the Senate Judiciary Committee from 2003 to 2010, that Thaler v. Perlmutter is 'the law of the land' and that human authorship is 'constitutionally required'. The Federal Judicial Center lists her Senate Judiciary service as 1993 to 2003 and her senior status from February 1, 2024; the D.C. Circuit (No. 23-5233, March 18, 2025) did not reach the constitutional question, and a denial of certiorari sets no precedent. 'Most important AI copyright ruling in history' and expertise predictions are removed. (1 page)
  • The Walker page said he has handled First Amendment, voting rights and government cases and that Florida has no circuit-level AI disclosure orders. No court source describes his docket; Florida Supreme Court AOSC26-12 (May 28, 2026, effective June 15, 2026) says courts may not impose AI disclosure or certification requirements by local order or other means. The page now also notes he was chief judge from 2018 to 2025 and that the local rules posted by the N.D. Fla. are the edition effective November 24, 2015. (1 page)

Source

Federal Judicial Center, Biographical Directory of Article III Federal Judges (judges.csv, read November 2, 2026): careers, senior-status and retirement dates; S.D.N.Y., Judge Broderick's Individual Rules and Practices in Civil Cases (revised October 29, 2025), section 4.J, read November 2, 2026; S.D.N.Y., Judge Caproni's chambers page and Individual Practices (revised March 17, 2025), read November 2, 2026; D.D.C., Judge Kelly's court page and Civil Standing Order template; Local Rules (September 2026), read November 2, 2026; D.D.C., Senior Judge Howell's court page and Standing Orders list, read November 2, 2026; D.C. Circuit, Thaler v. Perlmutter, No. 23-5233 (March 18, 2025), read November 2, 2026; N.D. Ill., Judge Kennelly's procedures page and the Fuentes, Johnston and Coleman AI provisions, read November 2, 2026; E.D.N.Y., Judge Chen's Individual Practices and Rules (2026), read November 2, 2026; Supreme Court, Murthy v. Missouri, No. 23-411 (June 26, 2024), read November 2, 2026; C.D. Cal., Judge Birotte's posted orders (JpsApi), read November 2, 2026; Florida Supreme Court, AOSC26-12 (May 28, 2026, effective June 15, 2026), read November 2, 2026; N.D. Fla., local rules and Judge Walker's court page, read November 2, 2026

Pages that carry this correction (9)

Judge pages (Kim, Sorokin, Burroughs, Gonzalez Rogers, Davila, Liman, Subramanian, Breyer, Nuffer): wording re-read against the courts' own orders and the Federal Judicial Center

· 9 pages

What changed

  • The Sorokin page said he was a federal public defender and an Assistant U.S. Attorney and that the bench's culture is moving toward AI transparency. The Federal Judicial Center lists him as Assistant Attorney General of Massachusetts (1994 to 1997), Assistant Federal Public Defender (1997 to 2005), magistrate judge (2005 to 2014) and district judge since 2014. His posted orders and the D. Mass. Local Rules (through September 8, 2026) do not mention AI, so the bench-culture and AI-views statements are removed. (1 page)
  • The Burroughs page said she is one of the most vocal federal judges on AI transparency, quoted her as wanting lawyers to disclose AI use and said she requires citation-check certifications. Her court page lists one standing order (electronic filing, March 12, 2026) and the Local Rules have no AI rule; the quotation and her views on AI in the judiciary come only from press coverage of a 2026 panel whose text was not available to us. They are removed. The Overjet v. VideaHealth description now follows press coverage (preliminary injunction denied July 2024, motion to dismiss denied March 2025) and no longer calls the annotations AI-generated. (1 page)
  • The Kim page said at least five N.D. Cal. judges have AI orders, that the district hears more technology litigation than any other court and that she carries a technology and IP caseload. The court's own page says only that some judges have AI disclosure requirements; her biography describes civil and criminal work with an emphasis on civil cases. Her Standing Order (effective October 7, 2025) and Patent Standing Order do not mention AI. The count, the ranking and the caseload description are removed. (1 page)
  • The Gonzalez Rogers page said she is presiding over Musk v. OpenAI and that the N.D. Cal. has the highest concentration of AI-related cases in the country. We did not verify the docket status or the ranking, so the page now attributes the first to press coverage and drops the second. Her civil (June 22, 2026), criminal (July 21, 2026) and patent standing orders contain no AI provision. (1 page)
  • The Davila page gave his senior status date as December 20, 2024; the Federal Judicial Center says December 11, 2024. The page also described what he held in the SB 976 social media rulings, called California the most aggressive state on AI legislation and described his docket as heavy in patent cases; none of this was read against a primary, so it is removed or attributed. His civil, patent and All Judges standing orders do not mention AI. (1 page)
  • The Breyer page called Justice Stephen Breyer "late". The Federal Judicial Center records his retirement on June 30, 2022 and no death, so the word had no source and is corrected to retired. The page also described Judge Breyer's view of AI and his Silicon Valley history without a source; that is removed. His general, civil jury trial and bench trial standing orders (2025) do not mention AI, and the page now says so. (1 page)
  • The Liman page said he was a Cleary Gottlieb partner, that the S.D.N.Y. is among the strictest environments for AI, and that the S.D.N.Y. ruled in February 2026 that AI-generated materials are not privileged. The Federal Judicial Center lists private practice in New York and Assistant U.S. Attorney service (1994 to 1999) with no firm. His civil and criminal Individual Practices contain no AI rule. The ruling is one judge's and is known to us only from press, so it is now attributed and its scope is not described. (1 page)
  • The Subramanian page said he was a partner at Susman Godfrey and presided over the Combs trial. The Federal Judicial Center lists private practice in New York from 2007 to 2023 with no firm or partner status, and we did not verify the Combs trial against the docket. Section 8.F of his civil practices (revised May 11, 2026) was re-read and is unchanged. (1 page)
  • The Nuffer page said he took senior status in 2019 and practiced in St. George for over 30 years. The Federal Judicial Center gives senior status on April 2, 2022 and private practice in St. George from 1978 to 2002. The bankruptcy trustee role, the prediction that Utah courts will adopt AI rules and the claim that Utah is one of the most AI-aware jurisdictions had no source and are removed. SB 149 is now described only as signed on March 13, 2024. (1 page)

Source

Federal Judicial Center, Biographical Directory of Article III Federal Judges (judges.csv, read November 2, 2026): careers, appointment and senior-status dates; D. Mass., Judge Sorokin's court page and posted orders, and Judge Burroughs's Standing Order re Electronic Filing (March 12, 2026), read November 2, 2026: no AI content; N.D. Cal., Magistrate Judge Kim's court page and standing orders (effective October 7, 2025), read November 2, 2026; N.D. Cal., Using AI tools in your case (court guidance: some judges have disclosure requirements), read November 2, 2026; N.D. Cal., Judge Gonzalez Rogers' Standing Order in Civil Cases, updated June 22, 2026, read November 2, 2026; N.D. Cal., Judge Davila's court page and standing orders (civil August 2022, patent), read November 2, 2026; N.D. Cal., Judge Breyer's General Standing Order, revised July 16, 2025, read November 2, 2026: no AI provision; S.D.N.Y., Judge Liman's chambers page and Individual Practices (civil effective April 7, 2025; criminal March 17, 2025), read November 2, 2026; S.D.N.Y., Judge Subramanian's Individual Practices in Civil Cases, revised May 11, 2026, section 8.F, read November 2, 2026; D. Utah, Judge Nuffer's chambers page and resource materials, read November 2, 2026, and Utah Legislature SB 149 bill status (signed March 13, 2024)

Pages that carry this correction (9)

Judge pages (Thompson, Otake, Hanks, Cooper, Koh, Curiel, Scarsi, Hwang, Martinez-Olguin): wording re-read against the courts' own orders and the Federal Judicial Center

· 10 pages

What changed

  • The Thompson page said we found no AI-specific standing order from her and that her orders impose no AI requirement. Her separate Standing Order on Civility and Professionalism (August 18, 2025), Section VII, covers AI: counsel must personally confirm the accuracy of AI research, and billing statements submitted for the Court's review must indicate AI usage. Her civil, criminal and patent orders have no AI text. The page now says so, and the N.D. Cal. district page no longer says her order has no AI provision without that qualification. (2 pages)
  • The Thompson page said she sits in the Oakland division and built a section on Oakland's docket. Her court page lists San Francisco, Courtroom 9, 19th floor, and her civil order names the San Francisco courthouse. The Oakland framing and the docket-mix claims are removed. (1 page)
  • The Otake page said she was a Hawaii state circuit court judge, was appointed in 2020 and is the first woman of Japanese descent on the court. The Federal Judicial Center lists her as a prosecutor and Assistant U.S. Attorney, confirmed August 1, 2018, with no state judgeship. The first-woman claim had no source and is removed. (1 page)
  • The Hanks page said General Order 2025-04 requires attorneys to thoroughly review AI-assisted filings, that Judge Hanks has served since 2014 and that practitioners report high standards in his chambers. The order cautions against filing AI-drafted papers unchecked and holds the signer responsible under Rule 11; the Federal Judicial Center lists him as confirmed April 20, 2015; the reputation and 'cases already seen' statements had no source and are removed. His posted Court Procedures contain no AI provision. (1 page)
  • The Scarsi page said the civil order we read is dated July 2026. The posted civil order is titled Civil Case Standing Order October 2026. Paragraph 15 is unchanged. (1 page)
  • The Hwang page said both orders carry the file date 09-25-26. The civil order was reissued with the file date 10-21-26 (dated October 21, 2026) and the same paragraph 5; the criminal order is dated September 25, 2026. (1 page)
  • The Martinez-Olguin page called the AI provision Section 4. It is paragraph 4 of Section H (Pleadings and Docketing) of her Civil Standing Order. The text is unchanged. (1 page)
  • The Cooper page said he is known for thorough case management and detailed opinions and that his chambers expects precisely cited briefing. The court's pages for him say nothing of the kind, so the sentences are removed. (1 page)
  • The Koh page stated predictions about how she would rule on AI questions, and superlatives about her record, as facts. Her record is unchanged and sourced; the predictions are now labelled as our inference and the superlatives are removed. (1 page)
  • The Curiel page gave 2025 as the effective date of the California AI Transparency Act (it became operative August 2, 2026 after AB 853 delayed it from January 1, 2026) and inferred what his courtroom requires of AI use from his career. The date is corrected and the inferences are removed. What his chambers rules say about AI was not re-checked in this pass. (1 page)

Source

N.D. Cal., Judge Thompson's Standing Order on Civility and Professionalism, dated August 18, 2025, Section VII (read November 2, 2026): AI use not prohibited, counsel must personally confirm AI research, billing statements submitted for the Court's review must indicate AI usage; N.D. Cal., Judge Thompson's court page (read November 2, 2026): location San Francisco, Courtroom 9, 19th floor; Federal Judicial Center, biographical directory of federal judges (judges.csv, read November 2, 2026): Otake nominated December 21, 2017, confirmed August 1, 2018, career as prosecutor; Hanks confirmed April 20, 2015; S.D. Tex. General Order 2025-04 (entered May 7, 2025, read November 2, 2026): cautions against submitting AI-drafted filings without checking them and holds the signer responsible under Rule 11; C.D. Cal. judges' procedures site (read November 2, 2026): Judge Scarsi's civil order is titled Civil Case Standing Order October 2026; Judge Hwang's civil order is dated October 21, 2026 and her criminal order September 25, 2026; N.D. Cal., Judge Martinez-Olguin's Civil Standing Order (August 26, 2026, read November 2, 2026): the AI provision is Section H, paragraph 4; D.D.C., Judge Cooper's page and the D.D.C. standing orders (read November 2, 2026): biography only, no statement about his case management

Pages that carry this correction (10)

Judge pages (Baylson, Johnston, Scholer, Young, Chutkan, Fitzgerald, Rosenberg, Casper, Lee, Newman, Durkin, Williams, Cannon): wording re-read against the courts' own orders and the Federal Judicial Center

· 13 pages

What changed

  • The Baylson page said his order demands specificity (which tool, what it did, how you verified), applies only to civil matters, is among the most detailed in the district, influenced colleagues and carries stated sanctions. His order (June 6, 2023) is one paragraph: a plain factual statement that AI was used in any way and a certification that every citation to the law or the record was verified, for attorneys and pro se parties in any paper filed. Tool, contribution, case type, sanctions and influence are not in it and are removed. (1 page)
  • The Johnston page said he was publicly vocal that AI orders are unworkable, issued his order because attorneys asked, and that Judge Martin B. McGee entered a similar order. His order is a short reminder that Rule 11(b), Rule 26(g) and any other relevant rule, including ethical rules, apply when AI is used, and cites a 2023 Judicature article. No motive or quotation supports the rest, and McGee is not an N.D. Ill. judge in the court's list or the Federal Judicial Center directory. Those statements are removed. (1 page)
  • The Scholer page said she was appointed in 2019 after being a partner at a major law firm, and that the N.D. Tex. has no district-wide AI rule. The Federal Judicial Center lists her confirmed March 5, 2018, with Dallas practice, the Texas 95th District bench and AAA arbitration and mediation. N.D. Tex. Local Civil Rule 7.2(f), effective September 1, 2026, requires generative AI disclosure in briefs. The page now says so. (2 pages)
  • The Chutkan page said she was a Federal Public Defender and a Boies Schiller partner, was randomly assigned the Trump case and consistently sentenced at or above DOJ recommendations, and that the DOGE case put AI governance in her courtroom. Her court biography says 11 years at the D.C. Public Defender Service and 12 years at Boies Schiller. The other statements had no primary source and are removed. The D.C. Courts AI policy order could not be read (dccourts.gov returned 403), so its contents are no longer described. (1 page)
  • The Fitzgerald page counted four or more C.D. Cal. judges with AI orders and described his docket by case type. The court's Judges' Procedures system lists AI provisions for four district judges (Hsu, Staton, Walter, Wright) and several magistrate judges, none for Judge Fitzgerald (his page and five posted orders have no AI text). No primary source describes his caseload, so the docket description is removed. (1 page)
  • The Rosenberg page said she chairs the Advisory Committee on Civil Rules and will shape federal AI rules. The Supreme Court's April 10, 2025 announcement says she chaired it from 2022 and would become Director of the Federal Judicial Center in August 2025; the committee's July 2026 roster lists Judge Sarah S. Vance as chair. Claims of complex civil litigation experience and ALI membership had no source and are removed. (1 page)
  • The Casper page said she prosecuted public corruption and that Judge Burroughs publicly advocated AI disclosure. The court's press release says she was Deputy Chief of the Organized Crime Drug Enforcement Task Force, her page lists her as Chief Judge since 2025, and her posted standing orders and the district's local rules have no AI text. No primary source shows the Burroughs advocacy, so it is removed. (1 page)
  • The Lee page described Concord v. Anthropic as an AI safety case. It is a copyright suit by music publishers over song lyrics (No. 5:24-cv-03811-EKL). The page now says so; a repeated paragraph was removed and Chhabria's case is named Kadrey v. Meta. (1 page)
  • The Newman page said Whaley v. Experian involved hallucinated citations and prompted his AI ban. In that order (November 16, 2023) the pro se plaintiff admitted using AI, the complaint was dismissed on pleading grounds and Judge Newman reminded the parties of Section VI of his Civil Standing Order, with no sanction and no mention of fabricated citations. The provision predates the case. (1 page)
  • The Durkin page omitted his senior status (December 26, 2023), ranked the district by filings, said three or more N.D. Ill. judges address AI and that judges are increasingly aligning with the Illinois Supreme Court policy. His posted Case Procedures have no AI text; two judges (Fuentes, Johnston) have AI text, and one points to the Illinois policy. The ranking and trend statements had no source and are removed. (1 page)
  • The Williams page predicted mandatory AI disclosure in Delaware, described his docket with statistics and said more than 60% of the Fortune 500 are incorporated there. None of his posted orders or the district's rules has AI text; the Delaware Division of Corporations says more than 66%. The prediction and docket statistics had no source and are removed. (1 page)
  • The Cannon page gave the Trump case assignment month, said she dismissed the case as unconstitutional and described her prosecutor work as major crimes and appellate matters. The dismissal order (ECF 672, July 15, 2024) dismissed the superseding indictment because the Special Counsel's appointment violated the Appointments Clause. The assignment month and the career description had no source and are removed, along with the career-ending and congressional-oversight wording. (1 page)
  • The Young page quoted his July 16, 2024 order as titled "In re: Local Civil Rule 7(E) Standing Order". The order is headed "In re: Local Civil Rule 7(E)" and "Standing Order"; the page now says so. His order has no AI text. (1 page)

Source

E.D. Pa., Judge Baylson's Standing Order re Artificial Intelligence, dated June 6, 2023 (read November 2, 2026): plain statement of AI use and certification that every citation was verified; N.D. Ill., Judge Johnston's judge page (read November 2, 2026): AI standing order reminding filers of Rule 11(b) and Rule 26(g); N.D. Tex., Local Civil Rules effective September 1, 2026, Rule 7.2(f): generative AI disclosure in briefs (read November 2, 2026); D.D.C., Judge Chutkan's court biography (read November 2, 2026): D.C. Public Defender Service, Boies Schiller; Supreme Court press release, April 10, 2025: Judge Rosenberg, chair of the Advisory Committee on Civil Rules since 2022, to become Director of the Federal Judicial Center; N.D. Cal., Judge Lee's Civil Standing Order, updated August 31, 2026 (read November 2, 2026), and the Concord v. Anthropic complaint (No. 5:24-cv-03811-EKL, Dkt. 337); S.D. Ohio, Judge Newman's Civil Standing Order, Section VI, and Whaley v. Experian, No. 3:22-cv-356, Doc. 26 (November 16, 2023); Federal Judicial Center, biographical directory of federal judges (judges.csv, read November 2, 2026): Scholer, Casper, Durkin, Williams, Cannon, Fitzgerald, Rosenberg

Pages that carry this correction (13)

Judge pages (O'Connor, Kronstadt, Donato, Woods, Connolly, Singhal): wording re-read against the courts' own orders

· 6 pages

What changed

  • The O'Connor page said Judge Kacsmaryk requires an AI certification with specific language about oaths and allegiance and that he followed Judge Starr. His requirements page asks for a certificate with the notice of appearance and a certificate that counsel read his requirements and understand Rule 11 responsibility. The oath and allegiance passage is his reasoning, not wording counsel certify, and we found nothing that dates his order after Judge Starr's. (1 page)
  • The Kronstadt page said the Central District of California has approximately four judges with individual AI standing orders. We read the posted orders of all 61 judges and magistrate judges listed and found at least nine with a generative AI provision. The page now says so, and says Judge Kronstadt's own posted orders contain none. (1 page)
  • The Kronstadt page said he participates in the court's Patent Pilot Program. We found no source for that. His posted civil package does include an Initial Standing Order for Patent Cases, which is what the page now says. (1 page)
  • The Donato page said he serves on the court's Local Rules Committee and may push for uniform AI rules. No court source we read says so, and the prediction that the district will formalize AI rules soon had no source, so both are removed. The page now records that his six posted standing orders contain no AI provision. (1 page)
  • The Woods page said his individual practices emphasize thorough preparation, accurate citations and compliance with scheduling orders, and that SDNY judges operate with an implicit expectation of AI transparency. His practices say nothing about citation verification or AI, and the district claims had no source. The page now describes what the practices cover and says Rule 11 is the operative standard. (1 page)
  • The Connolly page inferred an AI disclosure expectation from his litigation funding and ownership orders, and said his ownership order applies to every corporation. None of his eleven posted documents mentions AI, and the April 18, 2022 ownership order applies only to joint ventures, LLCs, partnerships and LLPs. The page now says that. (1 page)
  • The Singhal FAQ said the 11th and 17th Circuits withdrew their January 2026 orders. We found no source for the month, so it now says early 2026, as the body of the page already did. (1 page)

Source

N.D. Tex., Judge Kacsmaryk judge-specific requirements (read November 2, 2026): certificate with the notice of appearance on generative AI use, and a certificate that counsel read his requirements and understand Rule 11 responsibility; the oath and allegiance passage is his stated reasoning; C.D. Cal. judges' procedures site (apps.cacd.uscourts.gov, 61 judges, orders read November 2, 2026): at least nine judges have a generative AI provision; Judge Kronstadt's three posted documents have none; N.D. Cal., Judge Donato's page and six standing orders (read November 2, 2026): no AI provision and no mention of a committee seat; S.D.N.Y., Judge Woods's Individual Rules of Practice in Civil Cases, revised March 19, 2025 (read November 2, 2026): chambers communications, conferences, motions and trial procedures; no AI or citation-verification provision; D. Del., Chief Judge Connolly's chambers page and eleven posted documents (read November 2, 2026): the April 18, 2022 ownership order applies to joint ventures, LLCs, partnerships and LLPs; no document mentions AI; Florida Supreme Court AOSC26-12 (read November 2, 2026) says several circuits adopted AI orders 'over the past several months' and gives no month

Pages that carry this correction (6)

Voicemail share and first-to-call-back claim removed, because no primary source supports them

· 2 pages

What changed

  • The intake automation page said 35% of calls to law firms go to voicemail. We found no primary study for that share, only undated secondary summaries that disagree, so the figure is removed and the page says we have no verified figure. (1 page)
  • The personal injury page said the first firm to call back wins the client. No study we could open measures that, so the page now says prompt follow-up matters and that AI can make the first reply immediate. (1 page)

Source

ABA Journal, How do you handle phone calls from potential clients (search result, not opened as a primary here); secondary summaries of older intake call studies report 32% and about a third of calls reaching voicemail, undated, so no figure is quoted

Pages that carry this correction (2)

Unsourced 'hours to minutes' paraphrases ('half a day takes 45 minutes', 'work that used to take associates days, done in minutes', 'minutes versus days of manual review', 'multi-hour follow-up cycles') removed or rewritten without a duration on 15 practice, tool, comparison, guide and use-case pages

· 15 pages

What changed

  • This page said AI turned work that took a person half a day, days or hours into minutes (or vice versa: 'multi-hour follow-up cycles' replaced by real-time answers). No source measured those durations, so the sentence now says the work is done faster than by hand, or tells readers to time their own. (15 pages)

Source

Thomson Reuters Institute, Making the most of AI's potential time savings (September 12, 2024; page read November 2, 2026). It reports lawyers' own estimate of 4 hours a week saved from AI in general, a survey estimate and not a measurement; none of the removed task durations has a primary, so none was replaced by another number

Pages that carry this correction (15)

Gavel: profile and pricing page now name its owner, Relativity

· 2 pages

What changed

  • Relativity announced on June 12, 2026 that it had acquired Gavel. Both companies say customers keep the same product, team and support for now, and neither says whether the Gavel name or the plans will change, so the pages tell readers to confirm the plan and the contracting entity with Gavel. The pages previously described Gavel without naming its owner. (2 pages)

Source

Relativity press release (PR Newswire, June 12, 2026); Gavel, announcement on its own site (June 15, 2026)

Pages that carry this correction (2)

Twenty-three pages: unsourced 'at a fraction of the cost' savings claims removed

· 23 pages

What changed

  • Reworded claims that Claude, ChatGPT, Notion AI, MDR providers, Logikcull, DISCO, Everlaw or AI-first legal firms do the same work 'at a fraction of the cost' (tool profiles, comparisons, guides, one use case and the Crosby line). No price, source or independent benchmark supports a ratio; the wording now says we have no independent benchmark or drops the cost claim. The same sweep removed three figures from the same sentences that had no source: a $400,000-$800,000 annual in-house contract spend, a $20,000-100,000 yearly waste per firm on tools that basic automation could replace, and an e-discovery claim that 50 reviewers for 6 weeks became 10 reviewers for 2 weeks. (23 pages)

Source

Editorial rule, a cost claim needs a published price with vendor, URL and read date, or the page says there is no independent benchmark

Pages that carry this correction (23)

eImmigration review and Docketwise vs eImmigration: price, forms, AI and CRM statements corrected to the vendors' own pages, unsourced timings and form lists removed

· 2 pages

What changed

  • The comparison still said eImmigration shows no price. Its pricing page lists three plans, $55 to $110 per user per month depending on plan and billing period, so the page now gives them with the date read. (1 page)
  • The review said eImmigration supports 100+ USCIS forms including I-130, I-140, I-589 and N-400. Its pages say 300+ forms and name I-485, I-765 and I-131 only, so the page now says that and no longer lists the other forms. (2 pages)
  • The pages said eImmigration's AI is limited to summarizing USCIS memos, RFE notices and evidence, and that it has no comparable AI layer. Its AI page lists a Summarizer (it names AAO decisions), an Extractor, a Translator, AI-driven intakes and drafting through a Visalaw.ai integration; the pages now say so. (2 pages)
  • The comparison said Docketwise has no CRM and added AI deadline tracking and document checklists in 2025. Docketwise lists CRM and lead management and 8am IQ on its Pro plan and we found no source for the 2025 claim, so both statements were replaced with what it lists. (1 page)
  • Removed migration, setup and form-update timings for eImmigration and Docketwise, and the claim that eImmigration tracks government fees; neither vendor publishes them on the pages we read. (2 pages)

Source

eImmigration, pricing (Starter $60 / $55, Essentials $90 / $82, Complete $110 / $100 per user per month; 300+ forms; read 2026-11-02); eImmigration, AI tools (AI Workmates: Summarizer, Extractor, Translator; Visalaw.ai drafting; read 2026-11-02); eImmigration, forms (300+ forms; read 2026-11-02); Docketwise, pricing (Basic, Pro, Advanced; CRM/Lead Management and 8am IQ on Pro; read 2026-11-02)

Pages that carry this correction (2)

Eastern District of Virginia: 'fastest federal court in the country' claim withdrawn, replaced with uscourts.gov Table C-5 figures

· 2 pages

What changed

  • The E.D. Va. AI disclosure page called the Eastern District of Virginia the fastest federal court in the country and said trials happen sooner than in any other district. Table C-5 does not support that: its median time to trial is 4th of 43 districts reporting. The page now cites those figures and says fast, not the fastest. (1 page)
  • The West Virginia court disclosure page described Virginia's Eastern District as one of the fastest federal courts in the country. It now cites its Table C-5 rank (4th of 43 districts in median time to trial). (1 page)

Source

U.S. Courts, Federal Court Management Statistics, Table C-5 (12 months ending June 30, 2026): E.D. Va. median 20.2 months to trial (16 cases), 4th of 43 districts reporting; all civil cases 6.3 months, 21st of 94

Pages that carry this correction (2)

Eastern District of Virginia: 'faster than almost any other federal court' replaced with the Table C-5 rank

· 1 page

What changed

  • The E.D. Va. AI disclosure page said cases reach trial faster than almost any other federal court, an unsourced ranking. It now gives the Table C-5 figure: 4th of 43 districts reporting in median months to trial. (1 page)

Source

U.S. Courts, Federal Court Management Statistics, Table C-5 (12 months ending June 30, 2026): E.D. Va. median 20.2 months to trial (16 cases), 4th of 43 districts reporting

Pages that carry this correction (1)

Doctrine completion date set back to September 3, 2026, Shepard's and 'verified citations' claims reattributed, Lighthouse, Nuix and Exterro security and terms read

· 14 pages

What changed

  • These pages gave September 4, 2026 as the date RELX announced it had completed the acquisition of Doctrine, a date taken from a WebWire republication. The LexisNexis PressRoom release itself is dated September 3, 2026, so the pages now say September 3 and link to that release. This reverses the October 25 change that moved the date to September 4. (11 pages)
  • The Lexis+ AI and Westlaw AI comparisons with Claude said Lexis+ AI Shepardizes citations automatically and that Lexis+ AI and Westlaw AI provide or guarantee verified or real citations. LexisNexis describes Shepard's Citations as a way to check a citation's status and treatment, and a Stanford study found both vendors' research tools hallucinate between 17% and 33% of the time, so the pages now say the answers are linked to sources that still have to be read. (2 pages)
  • The Lighthouse, Nuix and Exterro comparison said we had read no security or terms pages. It now records what we read on November 2, 2026 for Lighthouse and Nuix, and states that Exterro's Trust Center needs JavaScript and was not read. (1 page)

Source

LexisNexis PressRoom, 'RELX Completes Acquisition of Doctrine and Sets Ambition to Serve One Million Legal Professionals Across Europe' (read November 2, 2026), dated September 3, 2026, 'today announced that it has completed the acquisition'; announced April 28, 2026. WebWire republished it dated September 4, 2026; LexisNexis, Lexis+ AI and Protégé product pages (read November 2, 2026), Shepard's Citations 'validate citations and confirm the status and treatment of authorities'; Shepard's Verify 'checks citations in AI responses, documents, and drafts'; Magesh et al., 'Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools' (Stanford RegLab and HAI, arxiv 2405.20362, full paper read November 2, 2026), Lexis+ AI and Westlaw AI-Assisted Research each hallucinate between 17% and 33% of the time; a linked citation can be real and still mislead; Nuix Solutions Terms, version of 1 July 2026, clause 1.10 (read November 2, 2026), Nuix agrees it will not use Customer Data to train any AI Systems; Lighthouse Master Spectra Subscription Agreement (read November 2, 2026), no use of Client Data or AI inputs and outputs to train AI services

Pages that carry this correction (14)

Judge Curiel page: his civil chambers rules do have an Artificial Intelligence section

· 2 pages

What changed

  • The Curiel page said he has not issued a specific AI disclosure standing order, while its own source file said his rules have an AI section. His civil chambers rules do: generative AI is not prohibited, but counsel and pro se parties must personally confirm the accuracy of AI-generated content, keep records of all prompts submitted to generative AI tools and certify personal verification in any submission containing AI-generated content, with sanctions for omitting the certification; a separate section requires notice of AI-generated evidence and bars AI-hallucinated assertions of fact. The page now says so. The posted PDF shows no revision date. (1 page)
  • The Curiel page advised disclosing AI use voluntarily if his chambers rules do not address AI. They do, so Step 5 now points to the certification and prompt-record duties. (1 page)
  • The Southern District of California page said no judge in the district has issued an individual standing order on AI. Judge Curiel's civil chambers rules have an AI section. The page now says so and says we have not checked the other judges in the district. (1 page)

Source

S.D. Cal., Judge Curiel's Civil Pretrial & Trial Procedures (civil chambers rules PDF, read November 2, 2026; the posted PDF shows no revision date): Artificial Intelligence section (personal verification of AI-generated content, records of all prompts, certification in any submission containing AI-generated content, sanctions for omitting it) and AI Generated Evidence section (notice to the opposing party, no AI-hallucinated assertions of fact)

Pages that carry this correction (2)

Claude, Harvey, Legora, CoCounsel and ChatGPT: performance comparisons without a published benchmark reworded or removed

· 10 pages

What changed

  • Removed the claim that ChatGPT Plus with good prompts 'matches Harvey's output quality' for '80% of legal AI tasks'. No published test compares the two; the Vals Legal AI Report evaluated Harvey Assistant but not ChatGPT. The page now says a general model may be enough for everyday tasks and that we have not measured it against Harvey. (1 page)
  • Removed 'Harvey wins decisively' (US litigation research) and 'Legora wins decisively' (European cross-border work), and the 'faster research, faster drafting, fewer associate hours' ROI line. We found no independent benchmark or timing data for either vendor; the page no longer ranks them on output quality. (1 page)
  • Removed the statements that Luminance handles multi-jurisdiction comparison better than Harvey, that Harvey outperforms on drafting and negotiation strategy, and that Harvey wins for research-heavy litigation. Neither vendor was compared in a published benchmark; the page now says to test both on your own documents. (1 page)
  • Removed 'CoCounsel wins on pure research quality; Harvey has the edge on multi-step workflows'. In the Vals Legal AI Report Harvey Assistant scored 94.8% against CoCounsel's 89.6% on Document Q&A and CoCounsel 77.2% against Harvey's 72.1% on Document Summarization; neither entered the research task. The page now quotes those figures with the read date. (1 page)
  • Removed speed and accuracy claims with no published test: that Harvey's templates produce brief drafts faster than Claude Opus 4.7, that CoCounsel's templates produce deposition outlines faster, that a citation-verification step gives 'the same end-state accuracy as Harvey's pre-validated cites', that 'no foundation model matches a vendor-integrated citation database', that Opus 4.7's xhigh effort 'wins' because Harvey is optimized for production drafting, and that CoCounsel's workflow 'beats' a Claude-plus-tooling build on large productions. The pages now say no independent benchmark exists and cite Stanford's 17% to 33% hallucination range for verification. (2 pages)
  • Replaced 'CoCounsel's citations ... more reliable than general AI tools' with what Stanford's study found: legal research tools hallucinate less than GPT-4 but still 17% to 33% of the time, and the study did not test CoCounsel itself. Replaced 'Harvey's breadth beats CoCounsel's research depth' with the Vals results (Harvey higher on three of the four shared tasks, CoCounsel on summarization, none on research). (1 page)
  • Removed the 90% and 70% thresholds for when to add Harvey, the claim that Claude plus manual Westlaw checks 'beats' CoCounsel for most firms and 'is faster', and the claim that CoCounsel, Claude or a drafting tool 'will outperform' MyCase IQ and Clio Duo. None had a published measurement; the pages now state the price difference or the fit and say no benchmark was found. (3 pages)

Source

Vals Legal AI Report (updated February 27, 2025; read November 2, 2026): Harvey Assistant, CoCounsel, Vincent AI and Oliver against a lawyer baseline. Claude, ChatGPT, Legora and Luminance were not evaluated; Magesh et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, Stanford (arXiv 2405.20362; read November 2, 2026): Lexis+ AI and Thomson Reuters research tools hallucinate 17% to 33% of the time, less than GPT-4

Pages that carry this correction (10)

Claude's 200K-token window, Opus 4.7 window comparisons and Claude 3.x leftovers replaced on compare, guides and AI tools pages

· 40 pages

What changed

  • Replaced the 200K-token window given for Claude. Anthropic's model docs list 1M tokens for its current models; the page now says so with the date we read it and that we have not checked the limit in the app. (33 pages)
  • Removed the 200K window given for Claude Opus 4.7 and the comparisons built on it against GPT-5.5 and Gemini. Anthropic's model docs do not cover Opus 4.7, so the page now says we have not verified its window. (8 pages)
  • Removed 'largest context window' wording for Claude. We have no source comparing vendors' windows, so the page no longer ranks them. (5 pages)
  • Removed 'Claude 3.7' and 'Claude 3.5 Sonnet' as the model named on the page, including the claim that Claude 3.7 outperforms Legalese Decoder in most tests, which we cannot source and have not tested. (4 pages)

Source

Anthropic, models overview (platform.claude.com, read November 2, 2026): current models Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, each with a 1M-token context window; Opus 4.7 is not covered

Pages that carry this correction (40)

The 200K-token Claude window replaced on 24 practice-area and best-of pages

· 25 pages

What changed

  • These pages gave Claude a 200K-token context window with no source. Anthropic's model docs list 1M tokens for its current models; the pages now say that, with the read date, and say we have not checked the limit in the Claude app. (25 pages)

Source

Anthropic, models overview (platform.claude.com, read November 2, 2026): current models Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, each with a 1M-token context window

Pages that carry this correction (25)

Claude Pro price now states its billing term on 30 pages: $20 billed monthly, $17 billed annually

· 30 pages

What changed

  • These pages gave Claude Pro as $20 a month without saying that this is the monthly-billing price. Anthropic's pricing page lists $17 a month with the annual discount ($200 billed up front) and $20 if billed monthly, so each price now carries its billing term. (30 pages)

Source

Anthropic pricing (claude.com/pricing, read November 2, 2026): Claude Pro $17 per month with the annual subscription discount ($200 billed up front), $20 if billed monthly; Team standard seat $20 per seat per month billed annually, $25 billed monthly

Pages that carry this correction (30)

Claude Team price now carries its billing term and read date; Claude Sonnet 3.5 no longer offered as a current option

· 9 pages

What changed

  • The Claude Team price of $25 per user per month is now stated as the monthly-billing price, with $20 for annual billing and the date we read Anthropic's pricing page. (7 pages)
  • Removed Claude Sonnet 3.5 as the middle-ground model and the claim that it matches Opus on drafting quality, which we cannot source. The page now lists the models Anthropic's docs show as current with their prices, and says we have not tested whether any matches Opus. (1 page)
  • Removed the first-person claim of having tested every Claude release since Sonnet 3.5; we hold no record of those tests. (1 page)

Source

Anthropic, Claude pricing (claude.com/pricing, read November 2, 2026; anthropic.com/pricing redirects there): Team standard seat $20 billed annually, $25 billed monthly, 2 to 150 seats; Anthropic, models overview (platform.claude.com, read November 2, 2026): current models Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5; no Claude 3.x entry

Pages that carry this correction (9)

Claude 3.7 Sonnet benchmark claims, the 200K-token window and 'one-quarter the price' replaced on 2 pages

· 2 pages

What changed

  • Two pages named Claude 3.7 Sonnet as the top-scoring model on legal benchmarks. We have no record of that benchmark and the model is no longer Anthropic's current line, so the pages now name the current models and say we have no current legal benchmark. (2 pages)
  • Replaced the 200K-token window given for Claude on two pages. Anthropic's model docs list 1M tokens for its current models; the pages say we have not checked the limit in the app. (2 pages)
  • Removed 'one-quarter the price' for Claude against Spellbook on the Spellbook vs Claude page. Spellbook publishes no price, so no ratio can be checked; the page now says Claude has a published price and Spellbook lists none. (1 page)

Source

Anthropic, models overview (platform.claude.com, read November 2, 2026): current models Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, each with a 1M-token context window; Anthropic, Claude pricing (claude.com, read November 2, 2026): Team seat prices; no Spellbook price is published anywhere we could read

Pages that carry this correction (2)

An unsourced '$5K-$15K annual' Claude API cost, and the all-in total built on it, removed from the Spellbook vs Claude Cowork comparison

· 1 page

What changed

  • The comparison gave an annual Claude API cost range at firm-scale usage and an all-in total for the Cowork path, both without a source. The pricing page gives only per-token rates and we found no published annual figure, so the page now says API fees scale with usage and that we have no measured annual figure, and gives the engineer cost without the total. (1 page)

Source

Anthropic pricing page (claude.com/pricing, read November 2, 2026) publishes per-token (MTok) rates and plan prices, with no annual API cost for a firm

Pages that carry this correction (1)

Claude '200K token context window' replaced with Anthropic's current figure on 42 use-case pages

· 42 pages

What changed

  • Replaced the 200K-token window given for Claude. Anthropic's model docs list 1M tokens for its current models; each page now carries the read date and says we have not checked the limit in the app. (42 pages)

Source

Anthropic, models overview (platform.claude.com, read November 2, 2026): current models Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5, each with a 1M-token context window

Pages that carry this correction (42)

GDPR Article 83(5) fine ceiling now cited to the Official Journal text, an unsourced 'months of manual document review' removed, and a 30-60 day review window and 'weeks apart' release cadence replaced by Anthropic's release dates

· 3 pages

What changed

  • The document redaction page said GDPR fines can reach 4% of worldwide annual turnover and that we had not re-verified the regulation text. We read Article 83(5) in the Official Journal text: up to EUR 20 million or, for an undertaking, up to 4% of total worldwide annual turnover, whichever is higher, for infringements such as the basic processing principles, data subject rights and transfers to third countries. The page now says so and the notice is removed. (1 page)
  • The regulatory investigations guide said internal investigations used to depend on months of manual document review. We found no published figure for how long manual review takes, so the page now says AI document review can help investigations reach preliminary findings sooner and that we have no measured figure. (1 page)
  • The firm AI policy template guide said new Claude versions typically take a 30-60 day committee review and that new models arrived weeks apart. We found no benchmark for the review window, so the page now leaves the window to the committee, and it dates the four releases from Anthropic's release notes (September 1, 22 and 28 and October 7, 2026). (1 page)

Source

Regulation (EU) 2016/679 (GDPR), Official Journal text, Article 83(5) (EU Publications Office, read November 2, 2026), sets fines up to EUR 20 million or, for an undertaking, up to 4% of total worldwide annual turnover of the preceding financial year, whichever is higher; Anthropic API release notes (read November 2, 2026) date Claude Fable 5.1 on September 1, Opus 5.5 on September 22, Sonnet 5.5 on September 28 and Haiku 5.5 on October 7, 2026

Pages that carry this correction (3)

Three unsourced figures: the Zaid 300% claim, BigLaw adoption statistics and a research-tool return projection

· 3 pages

What changed

  • The case valuation page said John K. Zaid & Associates reported a 300% increase in settlement offers on certain case types, with no link. The claim comes from an EvenUp press release and says 'up to 300%' in some cases with no case types named; the page now says so and links it. (1 page)
  • The future of BigLaw guide gave 2027 shares of work handled by AI, a 30% A&O Shearman efficiency gain, associates spending 40% more time on research, and 15-20% higher revenue per partner, without sources. The 2027 line is now labelled our projection, the A&O Shearman figure is attributed to Harvey's customer story (a vendor estimate for contract review), and the rest is removed. (1 page)
  • The AI legal research guide said a 50-attorney firm spending $15,000 a month should expect $45,000 to $75,000 a month in recovered capacity. No source supported it; it is replaced by break-even arithmetic with stated assumptions ($300 per attorney per month, $200 an hour, 1.5 hours). (1 page)

Source

EvenUp via Business Wire, February 27, 2026 (read via Wedbush InvestCenter copy, November 2, 2026): Zaid says offers rose 'up to 300%' in some cases, names no case types, example $10,000 or less to $30,000 policy limits; Harvey, A&O Shearman customer story (read November 2, 2026): ContractMatrix 'up to an estimated 7 hours' saved per contract review; no client satisfaction or revenue per lawyer figures

Pages that carry this correction (3)

Unmeasured '70-80%' shares and AI before/after durations removed, because no study or primary source measures them

· 29 pages

What changed

  • These pages gave a 70-80% share (of first-pass review work, of daily legal AI use cases, of a draft that is usable, of analyst hours or trials handled by Copilot, of the way to a final demand letter). No study or primary source measures those shares, so the pages now say there is no measured share. (11 pages)
  • These pages gave a before-and-after duration for AI work (for example 90 minutes instead of 4 hours, or review dropping from weeks to days) without a measurement or a primary source. They now say we have no measured figure and suggest timing your own work. (19 pages)
  • The page said a RAND study found document review is 70-80% of litigation costs. We have no primary source we could open for that share, so the page now says document review is a large part of e-discovery cost and quotes no share. (1 page)

Source

RAND Institute for Civil Justice, e-discovery cost study (page MG1208), the study the 70-80% litigation-cost share was attributed to; no share is quoted now

Pages that carry this correction (29)

Document-volume, custodian-count and state-count scenarios without a primary rewritten as illustrative wording on six pages

· 6 pages

What changed

  • This page gave document and custodian counts for a government investigation, and savings and tool costs tied to them. We found no study, guide or vendor report with a method that supports them, so the page now says a large investigation can involve millions of documents and many custodians, and that we have no measured figure. (1 page)
  • This page gave a custodian range and annual cost comparison for legal hold work. No source supported them, so they were removed and the page says we have no measured cost. (1 page)
  • This page said how many custodians a typical SEC investigation involves. No source supports a typical count, so the page now says an investigation can involve many custodians. (1 page)
  • This page gave a document-count range for class action discovery and a number of states for a pattern. No source supported either, so the page now says often millions of documents and many states. (1 page)
  • This page gave a state count and a number of regulatory frameworks for a data breach scenario. No source supported them, so the scenario now says many states and gives no count. (1 page)
  • This page gave a state count and analyst and outside counsel costs for a compliance monitoring scenario. No source supported them, so they were removed and the page says we have no measured cost. (1 page)

Source

Kroll Ontrack (ediscovery.com Pulse), Metrics Reveal Dichotomy in the Ediscovery Market, Business Wire, October 28, 2014 (read November 2, 2026). The only benchmark found: an average of 16 custodians per project in 2013 (65 in 2008), across all projects, so it supports none of the removed figures

Pages that carry this correction (6)

Case valuation page: Basich v. Microsoft '$41M-$61M', Lex Machina 'highest level in a decade' and EvenUp '$385 million / $2 billion' now attributed to their primary sources with link and date

· 1 page

What changed

  • The three figures were confirmed against their primary sources and are now attributed with link and date. The Basich range is a projection from Darrow's own model for a newly filed case, not a settlement or an independent estimate. Lex Machina's release says 2025 federal class action filings exceeded 12,200, the highest volume in the past decade. EvenUp's own announcement gives total capital raised of $385 million and a valuation over $2 billion; the page now says so as EvenUp's statement. (1 page)

Source

Darrow, Evaluating Litigation Risk: Basich v. Microsoft (February 16, 2026; read November 2, 2026): Darrow's own model estimates a likely settlement range of $41M to $61M, the 25th to 75th percentile of its simulations, for a newly filed case; no settlement is reported; Lex Machina 2026 Class Action Litigation Report, press release on LexisNexis (April 16, 2026; read November 2, 2026): 2025 federal class action filings above 12,200, the highest volume in the past decade; EvenUp, Series E announcement (October 7, 2025; read November 2, 2026): $150M round led by Bessemer, total capital raised $385 million, valuation over $2 billion

Pages that carry this correction (1)

Seven pages that told an unnamed 'one firm' or 'one company' anecdote with figures: removed because no named case or source supports them

· 7 pages

What changed

  • The Luminance entry said one mid-size firm reviewed 4,000 leases in 3 days and that a review that took associates two weeks now takes days. Neither had a named firm or a source, so both lines are removed and the page says we found no published benchmark. (1 page)
  • The international trade guide said one consumer goods company saved USD 12 million a year in duty. No company or source was named, so the line is removed and the page says we found no published case study. (1 page)
  • The litigation strategy guide said an Am Law 50 partner reported that Lex Machina data changed forum selection and improved outcomes. No person, firm or source was named, so the claim is removed. (1 page)
  • The data privacy guide said one enterprise client found 340 unmapped data stores. No client or source was named, so the figure is removed. (1 page)
  • The outside counsel spend guide said one surveyed general counsel moved USD 800K a year of contract review in-house for a USD 120K tool. No survey or source was named, so the figures are removed. (1 page)
  • The intellectual property page said a mid-size IP firm cut freedom-to-operate opinion turnaround from 3 weeks to 5 days. No firm or source was named, so the claim is removed. (1 page)
  • The future of BigLaw guide quoted an unnamed managing partner as telling the ABA Journal 'We tried to wait and see. We waited too long.' We could not find the quote, so it is removed. (1 page)

Source

Luminance homepage (read November 2, 2026) describes AI contract review for legal teams and customer quotes; it gives no 4,000-lease, 3-day review result; LexisNexis, Lex Machina (read November 2, 2026) describes analytics on the past behavior of parties, counsel, experts, judges and courts; it gives no forum-selection result from a named partner

Pages that carry this correction (7)

Anecdotes and hypothetical figures with no named matter or source removed or labelled illustrative; BlockFills Research Suite AI Dossier claim withdrawn

· 6 pages

What changed

  • This page said one firm found $180,000 in undisclosed cryptocurrency using Claude. It named no firm or matter and gave no source, and we found none, so the claim was removed and the page now says AI output on financial records is a lead for a forensic accountant to confirm. (1 page)
  • This page described a case where Claude found 147 unauthorized transactions totaling $340,000 in a vulnerable adult's bank records. It named no matter and gave no source, so it is now labelled an illustrative example, not a real case, without figures. (1 page)
  • This page gave examples with figures (a $450,000 demand settled for $180,000, 85,000 documents reviewed for $45,000, a 40% rise in case assignments, reports cut from a day to 3 hours) with no named matter or source. They are now labelled illustrative scenarios without figures. (1 page)
  • This page said one firm saw higher offers with EvenUp and another reported a large improvement. No firm or source was named and we found none, so the page now says we have not found an independent study or named firm result. (1 page)
  • This page wrote a hypothetical settlement range (USD 1.2 to 1.8 million, 70% of the time) and a 45% summary judgment rate as if they were data. The example is now labelled hypothetical and gives no figures. (1 page)
  • This page said Research Suite's AI Dossier feature analyzing docket entries in the BlockFills bankruptcy showed AI use in Delaware dockets. We found no source for an AI Dossier feature; the public BlockFills case page says only that its report was prepared with Research Suite by Stretto, so the page now cites that and says we could not verify the feature. (1 page)

Source

chapter11cases.com, BlockFills Parent Files Chapter 11 in Delaware (read November 2, 2026): case No. 26-10371 (TMH), filed March 15, 2026; says only that the article was prepared using Research Suite by Stretto, and does not mention an AI Dossier

Pages that carry this correction (6)

Real Examples with hours, dollar and percentage figures and no named matter or source relabelled as illustrative examples without figures on 10 practice-area pages

· 10 pages

What changed

  • This page described a Fortune 500 department that flagged a multi-million-dollar auto-renewal exposure across 28,000 contracts with Ironclad, and a boutique that shortened closings with more deals a year. It named no firm or matter and gave no source, so both are now labelled illustrative, not real cases, without figures. (1 page)
  • This page gave a Texas firm that cut intake-to-filing from 8 hours to 2.5 hours with 40% more cases, and an attorney who cut a first draft from 3 hours to 45 minutes. No firm or source was named, so both are now illustrative examples without figures. (1 page)
  • This page described a Massachusetts firm delivering Medicaid analyses in 24 hours instead of 2-3 weeks, a Florida attorney cutting guardianship reports from 4 hours to 45 minutes (80+ hours a year) and a firm with better satisfaction scores. None had a named matter or source; they are now illustrative examples without figures. (1 page)
  • This page said an attorney removed an AI-training clause worth potentially millions, an agency built AI riders across 50+ agreements in a day, and a licence fee came in at 3x the first offer. No matter or source was named, so they are now illustrative examples without figures or outcomes. (1 page)
  • This page said an attorney reviewed a 2,800-page Draft EIS in 3 days instead of 3 weeks, and a firm audited 8 clients for $40,000 to avoid $200,000+ in penalties. No matter or source was named, so both are now illustrative examples without figures. (1 page)
  • This page gave a Florida adoption workflow worth 640 hours a year, an Ohio attorney with 18 guardianships, and a Texas firm that files protective orders same-day. No firm or source was named, so they are now illustrative examples without figures. (1 page)
  • This page said an employment practices defense team uses Claude for case evaluations, without a named firm or source. It is now an illustrative example. (1 page)
  • This page gave legal aid anecdotes with figures (200 to 320 eviction cases on a $240 subscription, 8 hours to 2 for country-condition research, 2 hours to 30 minutes for protective orders, twenty organizations sharing templates). No organization or source was named, so they are now illustrative examples without figures. (1 page)
  • This page gave real estate anecdotes with figures (8 to 28 closings a month and +250% revenue, 340 leases reviewed in 4 days for $15,000 instead of $90,000, 15 versus 90 minutes per title commitment). No firm or source was named, so they are now illustrative examples without figures. (1 page)
  • This page described 47 depositions analysed with 23 contradictions found, 4 bad citations caught with Clearbrief, and a judge who granted summary judgment in 78% of med-mal cases but 31% with certain expert testimony. No matter or source was named, so they are now illustrative examples without figures. (1 page)

Source

Clearbrief homepage (read November 2, 2026) describes a cite-checking platform for citations to law and facts; it gives none of the 4-citation or overruled-case results this page had attributed to a defense team; LexisNexis, Lex Machina product page (read November 2, 2026) describes analytics on the behaviour of parties, counsel, experts, judges and courts; it gives none of the 78% and 31% summary judgment rates this page had attributed to a medical malpractice trial

Pages that carry this correction (10)

Two worked examples presented as recent real matters, with counts and review costs, removed from two guides

· 2 pages

What changed

  • This page described a recent antitrust class action with a document count, a first-pass result and a review cost compared with manual review. It named no matter and gave no source, and we found none, so the example was removed and the page now says we found no published benchmark of review savings on a class action. (1 page)
  • This page described a recent FCPA investigation where AI custodian identification added a number of custodians. It named no matter and gave no source, and we found none, so the example was removed. (1 page)

Source

Kroll Ontrack (ediscovery.com Pulse), Metrics Reveal Dichotomy in the Ediscovery Market, Business Wire, October 28, 2014 (read November 2, 2026). The only benchmark found: an average of 16 custodians per project in 2013 (65 in 2008), across all projects, so it supports none of the removed figures

Pages that carry this correction (2)

Entertainment page: unsourced Lex Machina IP-firm example relabelled as illustrative

· 1 page

What changed

  • This page said an IP litigation firm representing visual artists uses Lex Machina to track every AI copyright case and built a database of fair use decisions to choose favourable jurisdictions. It named no firm and gave no source, so it is now labelled an illustrative example, not a real case, with no result claimed. (1 page)

Source

LexisNexis, Lex Machina product page (read November 2, 2026) describes analytics on parties, counsel, judges and courts; it names no IP firm and no AI-copyright database built with it

Pages that carry this correction (1)

Practice-area pages: unsourced time-saving figures removed or labelled as our own arithmetic

· 28 pages

What changed

  • Removed the claim that AI does 80% of a prior art search in 20% of the time. No source gives that split; the page now says we found no published figure. (1 page)
  • Removed task-time comparisons (a task done by hand in hours versus minutes with AI) on practice-area pages. No source supported them; the pages now say AI can shorten the work and that we have no measured figure. (23 pages)
  • Removed an unsourced doubling claim for lead conversion, appeals capacity and regulatory volume, and an invented judge grant rate used as an example. (4 pages)
  • Relabelled worked ROI examples as illustrative arithmetic on stated assumptions, not measured results. (5 pages)

Source

Thomson Reuters Institute, Making the most of AI's potential time savings (September 12, 2024; page read November 2, 2026), where lawyers' own estimate is 4 hours a week saved from AI in general, with no per-task figure for any practice area

Pages that carry this correction (28)

Unsourced hours-saved and task-time figures removed or rewritten as formulas on 59 practice, tool, comparison and guide pages whose hours-claims exceptions were due to expire

· 59 pages

What changed

  • This page gave hours-saved and task-time figures for ai for patent lawyers. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai for trial lawyers. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai maritime admiralty law. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai mediation arbitration. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for employment labor. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for docketwise. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for harvey ai implementation cost beyond license. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for best ai for family law. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for best ai for solo practitioners. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai tools solo practitioners. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for claude design vs figma for legal tech prototyping. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai cost benefit analysis template. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai for data privacy compliance. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for annual ai audit checklist law firms. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for anthropic cowork legal plugin triage nda rapid review. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for best ai for ip lawyers. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for best ai for mid size law firms. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for claude design client portal mockups without vendor. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for claude design courtroom demonstratives evidence prep. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for claude design legal prototyping anthropic stack. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for claude design skill github legal tech builders. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for claude design system integration law firm brand. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for claude opus 4 7 api pricing vs 4 6 legal tech budgeting. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for claude opus 4 7 contract review. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for claude opus 4 7 legal drafting. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for claude word legal drafting workflow. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for how to audit your law firms ai use. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for how to train lawyers on ai. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for law firm ai policy enforcement guide. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for microsoft copilot excel billable hours analysis firms. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for microsoft copilot powerpoint trial presentations 2026. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai fintech legal. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai billing optimization corporate. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai brief writing ip litigation. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai brief writing lawyers. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai case valuation personal injury. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai contract drafting ip. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai contract drafting workflow. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai contract review employment. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai contract review real estate. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai deposition analysis. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai deposition prep corporate. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai deposition prep personal injury. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai discovery management criminal defense. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai discovery management ip. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai document automation templates. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai due diligence corporate ma. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai due diligence real estate. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai email drafting lawyers. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai financial analysis bankruptcy. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai legal project management. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai legal research family law. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai litigation analytics ip. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai litigation analytics personal injury. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai medical record analysis family law. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai medical record analysis personal injury. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai policy drafting corporate. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai policy drafting employment. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)
  • This page gave hours-saved and task-time figures for ai regulatory analysis real estate. No source supported them, so they were removed, replaced by a formula on the reader's own numbers, or labelled as our planning assumption. (1 page)

Source

Thomson Reuters Institute, Making the most of AI's potential time savings (September 12, 2024; page read November 2, 2026). It reports lawyers' own estimate of 4 hours a week saved from AI in general, a survey estimate and not a measurement; none of the removed task figures has a primary, so none was replaced by another number

Pages that carry this correction (59)

Six pages that said review used to need 'armies of contract attorneys' or that AI is 'faster than any human team': both absolutes removed because no source supports them

· 5 pages

What changed

  • The white-collar defense page said these firms have always needed armies of contract attorneys and months of review time, and that AI changed the math dramatically. We found no source for the headcount or the duration, so the page now says review has traditionally taken large teams, that we have no measured figure, and that AI can take on part of the first-pass review. (1 page)
  • The regulatory investigations guide said privilege logging required armies of contract attorneys and that pattern detection finds the key documents faster than any human team. We found no study comparing AI with human review teams on speed. The guide now says AI tools can help draft privilege logs and surface patterns for attorneys to check, and that we found no published comparison. (1 page)
  • The compliance auditing page said the privacy compliance surface is expanding faster than any human team can track. No source measures that, so the page now says the surface is expanding and tracking it by hand is getting harder. (1 page)
  • The M&A document review page said the work previously required armies of contract reviewers billing at $50-150/hour. We have no source for the headcount or the rate, so the page now says the work fell to contract review teams and that we have no measured figure. (1 page)
  • The DISCO Cecilia page promised more consistent quality than armies of contract reviewers working under deadline pressure. No measured comparison exists in what we read, so the page now lists faster time to production and lower review costs as DISCO's stated aims and says we have no measured comparison. (1 page)

Source

Grossman and Cormack, Technology-Assisted Review in E-Discovery Can Be More Effective and More Efficient Than Exhaustive Manual Review, Richmond Journal of Law & Technology 17(3), 2011 (abstract read November 2, 2026): TREC 2009 data show two technology-assisted teams beating a manual review of the whole collection on recall and precision; it gives no reviewer headcount and no result against 'any human team'

Pages that carry this correction (5)

Safety absolutes without a source: NotebookLM and Clearbrief 'cannot hallucinate' attributed or bounded, database-grounded research no longer said to 'eliminate' hallucination

· 21 pages

What changed

  • Two NotebookLM comparison pages and the vendor directory said NotebookLM cannot or will never hallucinate, has zero hallucination risk on uploaded documents or is inherently more reliable. Google's help page says it is designed to answer from your sources with inline citations and that it can make mistakes. The pages now say that and that it can still misread or misstate what is inside your files. (3 pages)
  • Eleven pages said database-grounded legal research (Westlaw AI, CoCounsel, Lexis+ AI) eliminates hallucination or will not hallucinate a case. The Stanford study found the LexisNexis and Thomson Reuters research tools hallucinated 17% to 33% of the time. The pages now say grounding reduces but does not remove the risk and cite the study. (11 pages)
  • Six pages stated as fact that Clearbrief cannot hallucinate. That is Clearbrief's own claim (clearbrief.com), with no independent test published. The pages now attribute it to Clearbrief with the date read and note that a checker can still misjudge whether a source supports a proposition. (6 pages)
  • The best AI for litigation page said to match each tool to the phase where it saves the most time, with no measurement behind it. It now says to match them to the phase where your firm spends the most hours and that we have no measured savings per phase. (1 page)

Source

Google NotebookLM help page (read November 2, 2026); designed to answer from the uploaded sources with inline citations, and 'can make mistakes'; Magesh et al., 'Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools' (Stanford RegLab and HAI, arxiv 2405.20362, abstract read November 2, 2026); the LexisNexis and Thomson Reuters research tools each hallucinate between 17% and 33% of the time, less than GPT-4 but not none; Clearbrief (clearbrief.com, read November 2, 2026); the vendor says its cite-checking is deterministic, not generative, and that it can't hallucinate; no independent test is published

Pages that carry this correction (21)

EvenUp ROI, 'wins your cases' and 'undeniable' claims without a primary reworded on 8 pages

· 8 pages

What changed

  • The EvenUp vs Supio vs Darrow page said one demand letter that raises a settlement by $5,000-$10,000 pays for months of EvenUp, named EvenUp the fastest ROI and said Supio saves the most hours; the Filevine vs EvenUp page said EvenUp wins your cases and that the per-demand ROI becomes undeniable. None of those had a primary: EvenUp's own site only markets higher settlements and an ROI calculator, and neither Supio nor Darrow publishes a verified return. The pages now say we found no independent study and tell readers how to test each tool on their own closed files. (2 pages)
  • Six pages used 'undeniable' for claims we could not source or that were our own editorial judgement: Harvey's efficiency gains, Darrow's results and a count of consumers affected, the national trend toward AI court rules (Middle District of Tennessee, Southern District of Alabama), the disruption of law firms (April 2026) and an in-house cost-per-matter argument. The wording was removed, softened to what we can support, or replaced with a note that no independent measurement was found. (6 pages)

Source

EvenUp, home page (evenuplaw.com, read November 1, 2026): markets '10x higher settlements', a 250K+ verdict and settlement dataset and an ROI calculator; no independent study cited; Darrow, home page (darrow.ai, read November 1, 2026): markets legal-exposure detection for organizations; no count of consumer claims or independent measure of results

Pages that carry this correction (8)

Harvey '50 million terms per week' removed from five more pages and unsourced 'most accurate' / 'best-in-class' claims reworded on six

· 11 pages

What changed

  • Five Harvey pages (use cases, vs Spellbook, is it worth it, alternatives, 2026 review) still said Harvey processes 50 million contract terms per week, and the use cases page called contract review its highest-volume use case and said firms reviewing 200+ contracts monthly see the clearest ROI. Harvey's own pages only say that teams extract 20M+ terms in review tables (blog, March 9, 2026, vendor-reported). The pages now attribute that figure to Harvey and say we found no published ranking of its use cases or threshold for return. (5 pages)
  • Six pages stated capability superlatives with no primary source: the research phase of law review writing as where AI 'saves the most' time, Claude producing 'the most accurate Bluebook citations', Everlaw's user experience as 'best-in-class' per competitors, Relativity's active learning as 'industry-leading', Midjourney and GPT Image 2 text accuracy as 'best-in-class', DocuSign integration as 'obviously best-in-class' and 'AI ROI undeniable' on the Delaware bankruptcy page. Each was reworded without the superlative, attributed to its vendor, or now tells the reader how to test it. (6 pages)

Source

Harvey, Introducing Agent Builder (harvey.ai/blog/introducing-agent-builder, March 9, 2026, read November 1, 2026): 'Teams use agents in Harvey to extract 20M+ terms in review tables'; vendor-reported, no weekly figure; Harvey, home page and blog index (harvey.ai, read November 1, 2026): no '50 million terms per week' and no ranking of use cases by volume

Pages that carry this correction (11)

EvenUp settlement and ROI claims attributed to the vendor, and Claude 'superior reasoning' claims replaced with what we could and could not find

· 3 pages

What changed

  • The EvenUp comparison said EvenUp generates demand letters that maximize settlement values, pays for itself on the first case and has an undeniable ROI, and the heading said it actually increases settlement values. EvenUp's own site only markets higher settlements, with a customer testimonial, and offers its own ROI calculator; we found no independent study of settlement outcomes or return. The page now attributes those figures to the vendor, says plainly that they are marketing and a testimonial, and tells readers to test the product on their own closed files. (1 page)
  • These pages said Claude answers with superior reasoning, or that its superior reasoning could make it better at complex financial analysis. We found no independent benchmark comparing Claude's legal reasoning with Lexis+ AI, or Claude with Copilot on complex Excel analysis. The pages now say that and suggest checking the output against sources or testing both tools on your own work. (2 pages)

Source

EvenUp, home page and product page (evenuplaw.com, read November 1, 2026): markets higher settlements, a customer testimonial, a 250K+ verdict and settlement dataset and an ROI calculator; no independent study cited

Pages that carry this correction (3)

Retired phrases found on ten more pages by replaying this log against every page: Copilot licensing totals, 'over 30 federal courts', Harvey training claims, a 70-80% split, Briefpoint 'purpose-built', Westlaw 'most accurate'

· 10 pages

What changed

  • Removed a count of federal courts with AI standing orders that two pages still gave after the count was retired: individual federal judges have issued such orders and we have not counted them. (2 pages)
  • Removed statements that Harvey's agents are trained on a firm's templates or on antitrust work product. Harvey says it never trains on customer data by default and builds a bespoke model for a firm only if the firm asks, and it publishes no training details. (2 pages)
  • Removed a 70-80% share of lawyers' work that Claude and ChatGPT were said to cover (no measured share exists) and 'purpose-built' as our own verdict on Briefpoint. (1 page)
  • Removed the claim that Harvey's legal-domain fine-tuning shows on complex analysis: Harvey publishes no training details and we have not tested that. (1 page)
  • Removed the callout that called Westlaw AI-Assisted Research the most accurate tool. The Stanford RegLab and HAI study of 2024 versions found Lexis+ AI accurate on 65% of queries and Westlaw AI-Assisted Research on 42%; the figures are old and CoCounsel and Vincent were not tested. (1 page)
  • Removed combined Copilot licensing totals of $39,600 a year for 50 attorneys that rested on Microsoft 365 base-license prices we could not re-read. The pages now give the $30 per user per month add-on alone ($18,000 a year for 50 attorneys) and point to Microsoft's current plans. (3 pages)

Source

Magesh et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (Stanford RegLab and HAI, first released May 2024); Harvey, security page (read October 25, 2026); Harvey, platform page (read October 25, 2026); Microsoft 365 Copilot, pricing (microsoft.com)

Pages that carry this correction (10)

Four sourcing fixes: an unsupported detail about an Alabama sanction removed, ChatGPT Plus and Microsoft 365 prices flagged as unconfirmed, and a California rule date corrected

· 39 pages

What changed

  • This page said the Alabama Supreme Court attorney sanctioned in April 2026 (W. Perry Hall, Ibach v. Stewart) apologized and then cited two more nonexistent cases in the apology footnote, and that he also owed double appeal costs. The report we could read (WFTV, April 28, 2026) supports the $17,200 in fees and costs, the State Bar referral and the bar on filing alone, but not those two details, so they were removed. (4 pages)
  • This page gave the ChatGPT Plus price as an exact figure. OpenAI's pricing page was not readable by our verifier, so the price is now given as about $20 a month, with a pointer to OpenAI's pricing page for the current figure. (27 pages)
  • This page listed Microsoft 365 list prices per user per month (Business Standard, Business Premium, E3, E5). Microsoft's pricing pages show prices only after a script runs, so the page now says the prices were not re-confirmed on November 1, 2026 and should be checked at microsoft.com. (7 pages)
  • This page said California Rule 10.430 was adopted in July 2025. The rule text on courts.ca.gov says 'adopted effective September 1, 2025', so the page now gives only that effective date. (2 pages)

Source

WFTV, Alabama Supreme Court sanctions, fines attorney over AI hallucinations in filings (April 28, 2026), reporting the court's opinion in Ibach v. Stewart; California Rules of Court, rule 10.430 (courts.ca.gov, read November 1, 2026): 'adopted effective September 1, 2025'

Pages that carry this correction (39)

Harvey agent and training wording sourced to Harvey's own pages, Clio SOC 2 note now cites Clio's Trust Centre, and the Kansas AI committee attributed to secondary reports

· 26 pages

What changed

  • These pages said Harvey's Agent Builder is trained on firm-specific material, that each firm's instance learns its institutional knowledge, and described Harvey as purpose-built or specific to a firm. Harvey's own pages say its Agent Builder lets a firm build custom agents from its own knowledge, playbooks and preferences, that it never trains on customer data by default, and that a bespoke model is built only if a firm asks. The pages now say that, with the date we read it, and label 'purpose built' as Harvey's own description. Comparisons about customization we had not tested were reworded as untested. (13 pages)
  • These pages said Clio's security documentation was not verified. Clio's Trust Centre (trust.clio.com) lists SOC 2 Type 2, so the pages now say that and leave whether the report covers Clio Duo and its AI data terms as a question for Clio. (12 pages)
  • The Kansas ethics page stated that the Kansas Supreme Court created an Ad Hoc Artificial Intelligence Committee in February 2025. The date rests on secondary reports of that order, so the page now attributes it to them and points readers to the order itself. (1 page)

Source

Harvey, Agents page (harvey.ai, read November 1, 2026): 'Purpose built agents'; Agent Builder: 'Build custom agents that capture your expertise'; Harvey, Security page (harvey.ai, read November 1, 2026): never trains on customer data by default; a bespoke model only on request; annual SOC 2 Type II and ISO 27001 audits; Clio Trust Centre (read November 1, 2026): lists SOC 2 Type 2 under Compliance

Pages that carry this correction (26)

Claude Team minimum corrected to 2 seats (range 2-150)

· 5 pages

What changed

  • These pages put the Claude Team minimum at five seats. Anthropic's pricing page, read on November 1, 2026, shows a Team minimum of 2 seats, so the pages now say 2-150 seats and cite the page. (5 pages)

Source

Claude pricing (claude.com/pricing, read November 1, 2026): Team plan minimum of 2 seats

Pages that carry this correction (5)

Harvey '50M terms/week' replaced and generic 'saves the most' claims reworded on 10 pages

· 10 pages

What changed

  • The Harvey use cases page said Harvey handles contract review at 50M terms per week. Harvey's own post (March 9, 2026) says teams extract 20M+ terms in review tables and has no 50M weekly figure, so the page now quotes the 20M+ figure as vendor-reported. (1 page)
  • Nine pages said that a given task, deliverable or workflow 'saves the most' time or money (M&A due diligence, page-line deposition digests, regulatory client communications, data room indexing, inconsistency identification, non-compete analysis, one contract drafting workflow, litigation document review, workflow learning); the litigation page also said Everlaw reviews documents 'better than anyone'. We found no source that measures any of these, so the sentences now describe the task without ranking it. (9 pages)

Source

Harvey, 'Introducing Agent Builder' (harvey.ai, post of March 9, 2026, read November 1, 2026): teams use agents in Harvey to extract 20M+ terms in review tables; the post does not mention 50M terms per week (vendor-reported)

Pages that carry this correction (10)

Unsourced hours-saved figures and the USD derived from them removed or rewritten as formulas on the last 17 pages of the lote 1 inventory

· 17 pages

What changed

  • This page gave 3-5 hours saved per attorney per month on billing narratives (worth $4,500-7,500 for five attorneys) and 5-8 paralegal hours a month on settlement distributions. No source supported them; the dollar example is now labelled arithmetic on an assumed 4 hours and $300/hour. (1 page)
  • This page gave hours for employment case valuation (8-15 manual, 3-5 with AI), EEOC charge evaluation, case screening, 40-60 hours a month and $3,000-7,500 a case. No source supported them; the monthly figure is now labelled arithmetic on an assumed 2 hours per case and the dollar figure was removed. (1 page)
  • This page said spreadsheet work eats 3-5 hours a week per attorney and that a monthly billing report now takes 20 minutes. We have no measurement for either, so both were removed. (1 page)
  • This page said Claude saves 3-4 hours per motion and NotebookLM 1-2 hours of preparation per case. No source supported them, so they were removed. (1 page)
  • This page gave hours for immigration regulatory monitoring, policy impact assessment and client updates, manual versus AI. No source supported them, so they were removed. (1 page)
  • This page said prior-art search preparation drops from 8-12 hours to 2-3 hours and that the tools save a team 15-25 research hours a week. No source supported them, so they were removed. (1 page)
  • This page gave hours for PI demand letters, discovery responses and motions, manual versus AI, and 20-35 hours a week saved across a practice. No source supported them, so they were removed. (1 page)
  • This page gave hours for real estate compliance monitoring, impact memos and audit preparation, manual versus AI. No source supported them, so they were removed. (1 page)
  • This page gave hours for employment compliance monitoring, advisory memos and handbook audits, manual versus AI, and 15-20 hours a month saved. No source supported them; the monthly figure is now labelled arithmetic on an assumed 2 hours per memo. (1 page)
  • This page said a family-law attorney spends 8-12 hours a week on client communication, that AI cuts it to 3-5 hours and saves 25-30 hours a month worth $7,500-9,000. No source supported them, so they were removed. (1 page)
  • This page said a three-person PI firm in Texas saves 5-8 hours a week with Custom GPTs. We could not verify it, so the figure was removed. (1 page)
  • This page said AI saves 15-25 hours on a 30-page law review article, with per-phase figures of 5-8, 8-12 and 3-5 hours. No source supported them, so they were removed. (2 pages)
  • This page gave hours for immigration form preparation (2-3 hours saved a case), RFE responses (6-8 hours to 2-3) and data entry. No source supported them, so they were removed. (1 page)
  • This page said AI saves 10-15 hours per probate estate and that a workflow takes 30-60 minutes instead of 3-5 hours. No source supported them; the dollar example is now labelled arithmetic on an assumed 10 hours per estate. (1 page)
  • This page used 10-20 hours a month of AI use as a fact about light users. It is now labelled an assumption. (1 page)
  • This page said solo practitioners report saving 10-15 hours a week on research and drafting. No source supported it, so it was removed. (1 page)
  • This page gave 5-10 hours a week of an implementation champion's time as a budget. It was our planning suggestion, not a measured figure, and is now labelled so. (1 page)

Source

Thomson Reuters Institute, Making the most of AI's potential time savings (September 12, 2024), read November 1, 2026; lawyers' own estimate of 4 hours a week saved from AI in general, a survey estimate and not a measurement; none of the removed task figures has a primary, so none was replaced by a number

Pages that carry this correction (17)

Copilot and AI-search guides: schema markup is no longer presented as a proven way to get cited by AI

· 5 pages

What changed

  • This page said that schema markup (FAQPage, Article, Person) is what AI grounding rewards, or that pages without it are invisible to AI. We found no primary source for that. Google says no special schema.org markup is needed for AI Overviews or AI Mode, and we found no Microsoft statement that Copilot requires schema. Bing's own AI Performance guidance recommends clear headings, tables and FAQ sections, which is a content point, not a markup one. The page now says we have not isolated the effect of markup on our own citations and attributes what the primaries say. (5 pages)

Source

Google Search Central, AI features and your website (updated December 10, 2025; read October 31, 2026); Google Search Central, General structured data guidelines (updated July 10, 2026; read October 31, 2026); Bing Webmaster Blog, Introducing AI Performance in Bing Webmaster Tools (February 2026; read October 31, 2026)

Pages that carry this correction (5)

MyCase IQ vs Clio Duo: MyCase Pro and Advanced prices now match MyCase's own pricing page

· 1 page

What changed

  • This page gave MyCase Pro as $89 and Advanced as $109 per user per month on annual billing, called Basic $39 and said monthly billing adds $10 on every tier. MyCase's pricing page, read October 31, 2026, lists Basic at $50, Pro at $100 and Advanced at $130 per user per month billed annually, and $60, $120 and $150 billed monthly. The page now shows those figures, which also match our MyCase, Smokeball and Clio Duo comparison. (1 page)

Source

MyCase pricing page (read October 31, 2026)

Pages that carry this correction (1)

District of Colorado page: Judge Crews is a District Judge and Judge Neureiter's AI order dates from September 22, 2026

· 1 page

What changed

  • This page called Judge S. Kato Crews a magistrate judge and dated the Colorado certification orders "as of April 2026". The court lists Judge Crews as an Article III District Judge, and Magistrate Judge N. Reid Neureiter's standing order took effect September 22, 2026. The page now names each judge's correct title, the order dates and each order's scope. (1 page)

Source

D. Colo., Magistrate Judge N. Reid Neureiter, Standing Order Regarding the Use of Generative AI in Court Filings, effective September 22, 2026 (read October 31, 2026); D. Colo., District Judge S. Kato Crews, Standing Order for Civil Cases, revised January 1, 2026 (read October 31, 2026)

Pages that carry this correction (1)

Clio and Smokeball prices: figures with no readable primary source removed from four pages

· 4 pages

What changed

  • This page gave Smokeball a flat per-user monthly price. Smokeball's pricing page, read October 31, 2026, publishes no prices and asks firms to request a quote. The page now says so. (1 page)
  • This page priced the core elder law stack with a Clio Duo monthly price and said Clio Duo is included on Clio's Core plan and above. We have no checked Clio source for either claim as of October 31, 2026. The page now points to Clio's pricing page and gives no Clio number. (1 page)
  • This page's budget range, litigation total and daily cost included a Clio monthly price with no checked source as of October 31, 2026. The page now totals only the tools with a published price (Claude Pro, ChatGPT Plus and Briefpoint: 40 dollars a month, or 190 with Briefpoint) and names Clio without a figure. (1 page)
  • This page's comparison table gave Clio a monthly per-user price range. Clio's per-user price was not confirmed on its pricing page on October 31, 2026, so we have no checked Clio price. The table now points to Clio's pricing page and gives no Clio number. (1 page)

Source

Smokeball pricing page (read October 31, 2026; four plans, no prices, quote only); Clio pricing page (where to check Clio's current prices)

Pages that carry this correction (4)

Unsourced hours-saved figures and the USD derived from them removed or rewritten as formulas on 15 more pages, including the Microsoft Copilot cost model

· 15 pages

What changed

  • This page said the automation layer saves 10-15 hours a week per attorney and the AI layer 15-25 hours a week. No source supported either figure, so the page now says we have no measured figure. (1 page)
  • This page gave before-and-after drafting hours for a standard estate plan, complex trusts and intake memos, and 25-40 hours a month for a small practice. No source supported them; the monthly figure is now labelled arithmetic on an assumed 2 hours per plan. (1 page)
  • This page said that a paralegal task takes 4-6 hours a case, that drafting takes 3-4 hours, or that firms save 20-30 hours a month (personal injury and construction law practice pages). No source supported them, so they were removed. (2 pages)
  • This page gave before-and-after hours for RFE responses and asylum briefs, hours saved per case on translation, and 40-60 hours a month for a practice. No source supported them; the monthly figure is now labelled arithmetic on an assumed 2 hours per case. (1 page)
  • This page gave hours for bankruptcy research (exemption comparisons, plan confirmation, Chapter 11) and a monthly saving of 30-45 hours worth $7,500-11,250. No source supported them; the monthly figure is now labelled arithmetic on an assumed 1 hour per research issue. (1 page)
  • This page gave monthly hours saved per attorney for five practice areas and hours per task, as 'realistic benchmarks'. No source supported them, so the page now lists the tasks to time and shows the break-even as arithmetic on a billing rate. (1 page)
  • This page said case-law research drops from 4-6 hours to 45-90 minutes, theory development saves 1-2 hours a case and a five-attorney firm saves 15-25 hours a week. No source supported them, so they were removed. (1 page)
  • This page gave adoption-preparation hours, emergency-filing hours and 5-15 hours a week saved by document automation. We could not verify them, so they were removed. (1 page)
  • This page gave hours for closing packages, purchase agreements and commercial leases, and 35-65 hours a month saved. No source supported them; the monthly figure is now labelled arithmetic on an assumed 2 hours per closing. (1 page)
  • This page said that AI saves 5-10 hours a month, that keeping up takes 2-3 hours a week, or that small habits save 2-3 hours a day (court-deadline practice page, legal AI podcast guide, productivity-hacks guide). None was measured, so they were removed or marked as our suggestion. (3 pages)
  • This page gave hours for criminal-defense motions, sentencing memoranda and appellate briefs, manual and AI-assisted. No source supported them, so they were removed; the dollar example is labelled arithmetic on an assumed 5 hours per motion. (1 page)
  • This page derived USD from hours nobody measured: 300-650 configuration hours worth $45,000-$140,000, a year-one cost of $63,000-$158,000, $1.1M-$2.25M a year of recovered time and a breakeven in 2-6 weeks. None has a primary source. The cost model is now a formula on figures the reader supplies (hours x rate) plus arithmetic on a labelled assumption; the only published figure kept is the $30 per user per month price, read on Microsoft's pricing page on October 31, 2026. (1 page)

Source

Thomson Reuters Institute, Making the most of AI's potential time savings (September 12, 2024), read October 31, 2026; lawyers' own estimate of 4 hours a week saved from AI in general, a survey estimate and not a measurement; none of the removed task figures has a primary, so none was replaced by a number; Microsoft 365 Copilot pricing, read October 31, 2026; $30.00 per user per month, paid yearly (the only published figure the Copilot page keeps; the other published prices keep their October 24, 2026 check dates)

Pages that carry this correction (15)

Unsourced 'saves N-M hours' and per-task time figures removed or labelled as assumptions on 23 more pages

· 23 pages

What changed

  • This page said (like ten other practice-area pages: banking, veterans, franchise, aviation, agricultural, election, food and beverage, municipal, nonprofit governance, space and telecommunications law) that AI saves 5-10 hours a month on the minimum stack, 20-30 hours a month on CoCounsel and 5-15 hours a month overall. No source supported those figures, so the pages now say we have not measured them, and the monthly dollar return is labelled as arithmetic on an assumed 5 hours. (11 pages)
  • This page said Gavel saves 1-2 hours per closing and put a monthly dollar value on it. The hours had no source; the dollar value is now labelled as arithmetic on an assumed 1 hour per closing. (1 page)
  • This page said PI firms report EvenUp saves 10-15 hours per case (market-landscape guide) or that Supio saves 3-5 hours per case (EvenUp vs Supio vs Darrow). Neither vendor publishes such a figure that we could verify, so both were removed. (2 pages)
  • This page gave before-and-after hours for verdict research (3-5 hours to 20-30 minutes), medical-record review (4-6 hours to 1-2 hours) and demand letters (2-3 days to 4-6 hours), and 30-40 hours a month recovered. No source supported them; the page now says we have no measured figure and labels the monthly capacity as an illustration. (1 page)
  • This page said document-universe mapping saves 200-400 hours on a mid-market deal. No source supported it, so the figure was removed. (1 page)
  • This page said (Visalaw, Imagility and Docketwise pages) that a petition or case takes 2-4, 1-2, 2-3, 3-5 or 4-6 hours depending on the tool. The vendors publish no such figures that we could verify, so they were removed. (3 pages)
  • This page said Briefpoint cuts a 2-4 hour task to 15-30 minutes and that paralegals save 5-15 hours a week. No primary source supported either, so both were removed. (1 page)
  • This page said leading firms' attorneys save 10-15 hours a week and middle-of-the-pack firms 2-5 hours. The only dated primary (Thomson Reuters Institute, September 12, 2024) is a survey estimate with a different figure, so the numbers were removed. (1 page)
  • This page gave 2-4 hours a month of maintenance per internal tool (two Claude Design guides). It was an unmeasured estimate, so the pages now say we have no measured figure. (2 pages)

Source

Thomson Reuters Institute, Making the most of AI's potential time savings (September 12, 2024), the dated primary on lawyer time savings (a survey estimate, not a measurement); none of the removed hours figures has a primary, so none was replaced by a number

Pages that carry this correction (23)

Unsourced senior-counsel and 'partner time' hours removed on 2 more Anthropic guides

· 2 pages

What changed

  • This page said configuring the Anthropic legal-skills playbook takes 4-15 hours of senior-counsel time (4-6 hours for a small firm, 8-15 for a mid-market firm), 2-3 hours per quarter of governance and 1-2 hours a month of upstream monitoring. Anthropic's plugin pages carry no hours figure and we have no measured one, so the figures are removed and readers are told the time depends on scope. (1 page)
  • This page said the playbook takes a partner 4-6 hours to write and 12 hours of partner time in total for a 25-attorney deployment. No source supported either figure, so the page now says we have not measured the partner time. (1 page)

Source

Thomson Reuters Institute, Making the most of AI's potential time savings (September 12, 2024), the dated primary on lawyer time savings; none of the removed hours figures has a primary, so none was replaced by a number

Pages that carry this correction (2)

Unsourced 'saves N-M hours' and 'partner time' claims removed or labelled as assumptions on 12 more pages

· 12 pages

What changed

  • This page gave the Anthropic legal-skills playbook as costing 4-15 hours of senior-counsel setup and 2-3 hours per quarter of review, 4-6 hours of partner time for a 25-seat firm, and 12-17 hours a month recovered on retainer reviews. Anthropic's plugin pages carry no hours figure and we have no measured one. The figures are removed, and the retainer example is now arithmetic on a visible 10-minutes-per-agreement assumption. (1 page)
  • This page said lawyers report saving 2-4 hours per research assignment with Harvey and 15-30 minutes a day on billing descriptions with Clio Duo. Neither had a source. Both are removed, and the monthly dollar examples are now labelled arithmetic on 22 working days. (1 page)
  • This page said Claude cuts work on a major authority from 4-6 hours to 1-2, that Clearbrief cuts citation verification from 4-6 hours to 2-4, and that a 2,000-page record takes Claude 2-3 hours. No source supported any of them, so the page now tells readers to time their own. (1 page)
  • This page said CoCounsel's break-even was 1-2 saved hours a month on Westlaw and 3-6+ without it, with a 5 and 15 hours-a-month rule of thumb. The 1-hour figure is now shown as the cited $340.85 price divided by a $350 hourly rate; the rest depends on quotes the page does not have, so it now says to divide your own quote by your hourly rate. (1 page)
  • This page said lawyers need 10-20 hours of training to become proficient with legal AI, and that saving 2 hours a week is worth $3,000-$4,000 a month. The training figure had no source and is removed; the dollar figure is now labelled arithmetic on a 4.33-week month and a $350-$450 rate. (1 page)
  • This page said that a Harvey user who saves 1-2 hours a month loses the firm money. The hours figure had no source; the page now says only that low adoption loses money and tells readers to run their own adoption numbers. (1 page)
  • This page said the Team Premium upgrade pays back if the user gains 4-6+ hours a week. No source supported it, so the page now tells readers to divide the price differential by the user's hourly rate. (1 page)
  • This page said a Gavel template takes 2-5 hours to build, with a simple one taking an hour and a complex trust a full day. Gavel publishes no such figure. The page now says to time the first template and shows the break-even as arithmetic on a 5-hour build. (1 page)
  • This page said AI cuts trust administration from 15-20 hours a year to 5-8, that WealthCounsel generates a trust package in 20-30 minutes against 3-4 hours by hand, that it costs $175-350 a month, and that the hybrid workflow cuts a 4-hour plan to 90 minutes. None had a source (WealthCounsel's price is by quote), so the figures are removed and readers are told to time their own plans. (1 page)
  • This page said a legal memo takes a junior associate 10-20 hours and an AI-assisted workflow 3-7. No source supported either range, so the page now says we have no measured hours and tells readers to time a few memos both ways. (1 page)
  • This page said AI cuts high-asset divorce financial review from 40-80 hours to 10-20, hidden-asset screening from 15-25 hours to 3-6, and saves 200-400 hours and $50,000-$140,000 a month. No source supported them. The page now shows a labelled illustration (1 hour saved per case) and tells readers to use their own measured figure. (1 page)
  • This page said AI cuts Chapter 7 petition preparation from 6-10 hours to 2.5-4, Chapter 13 plan drafting from 3-5 hours to 1-2, and saves 50-80 hours a month. No source supported them. The page now tells readers to time their own filings and shows a labelled illustration (2 hours saved per filing). (1 page)

Source

Thomson Reuters Institute, Making the most of AI's potential time savings (September 12, 2024), the dated primary on lawyer time savings; none of the removed hours figures has a primary, so none was replaced by a number

Pages that carry this correction (12)

Unsourced 'saves N-M hours' claims removed or labelled as assumptions on 12 pages

· 12 pages

What changed

  • This page said MedChron and Supio each save 5-10 hours per case on record review. We have no independently measured figure for either tool, so the page now tells readers to time a few of their own record reviews. (1 page)
  • This page said firms using AI save 5-10 hours per attorney per week, worth $6,000-12,000 a month. No primary source supported it. The page now cites the Thomson Reuters Institute's September 12, 2024 write-up of its 2024 Future of Professionals survey, in which lawyers estimated an average of 4 hours saved per week in the first year (an estimate, not a measurement), and shows the dollar figure as labelled arithmetic. (1 page)
  • This page gave family-law document preparation as falling from 4-6 hours to 1-2 hours and a 30-50 hour monthly saving. We have no measured figure, so the claim is removed and the monthly figure is now labelled arithmetic on a 1-hour assumption. (1 page)
  • This page said 25 interrogatories drop from 4-6 hours to 1-2 and that most litigators save 2-4 hours per set of discovery responses. Neither had a source. Both are removed and readers are told to time their own sets. (1 page)
  • This page said most research-intensive attorneys save 10-20 hours with CoCounsel. No source supported it. The page now shows the break-even as labelled arithmetic and says we have not measured the saving. (1 page)
  • This page said a MedChron chronology saves 3-5 hours per case and a DemandsAI draft 2-4 hours. We found no independently verified figure, so the claim is removed and readers are told to time their own cases. (1 page)
  • This page said firms doing 30+ closings a month save 2-3 hours per transaction, worth $15,000-20,000 a month. No source supported it. The page now shows the arithmetic on a 1-hour assumption and says to replace it with a timed figure. (1 page)
  • This page said automating NDA and vendor agreement workflows frees up 15-20 hours per week for a typical in-house team. No source supported it, so the figure is removed. (1 page)
  • This page said DISCO's integrated approach saves a 5-attorney team 5-10 hours per week in context-switching. No source supported it, so the figure is removed. (1 page)
  • This page said CoCounsel is justified if each attorney saves 3-5 hours of research time per month. The threshold had no source, so the page now tells readers to track their own research hours for a month. (1 page)
  • This page said paralegals reclaim 8-10 hours per week within a month. No source supported it, so the figure is removed and readers are told to time each tool. (1 page)
  • This page said PI firms spend 20-40 hours per case organizing medical records manually. No source supported it, so the figure is removed. (1 page)

Source

Thomson Reuters Institute, Making the most of AI's potential time savings (September 12, 2024)

Pages that carry this correction (12)

Thomson Reuters Regulatory Intelligence is now CUBE: two compliance pages no longer present it as a current Thomson Reuters product

· 2 pages

What changed

  • The AI compliance monitoring guide and the AI compliance auditing use case described Thomson Reuters Regulatory Intelligence as a current Thomson Reuters product. CUBE announced on May 16, 2024 that it was buying the Thomson Reuters Regulatory Intelligence and Oden businesses and said it completed the purchase on December 31, 2024, so both pages now say CUBE (formerly Thomson Reuters Regulatory Intelligence). The '1,000+ regulatory bodies', '$500/month per user', '$500-1,500/month per user' and 'banks and broker-dealers rely on it' statements had no source and were removed. (2 pages)

Source

Hg Capital, announcement of May 16, 2024 (CUBE and the Regulatory Intelligence and Oden businesses); CUBE completion notice, January 2, 2025 (AAP release)

Pages that carry this correction (2)

Judge Richard Andrews page no longer says he presided over Thomson Reuters v. ROSS; the AI legal research guide no longer lists ROSS Intelligence's successor tools

· 2 pages

What changed

  • The Judge Richard Andrews page said he presided over Thomson Reuters v. ROSS Intelligence and described his AI expectations from that case. The Third Circuit's opinion names Judge Stephanos Bibas, a Third Circuit judge sitting by designation, as the district judge and does not mention Judge Andrews. The page now says so, cites no AI standing order from Judge Andrews, and drops the claims about his AI knowledge and disclosure expectations. (1 page)
  • The AI legal research guide listed "ROSS Intelligence's successor tools" among the tools carving out niches. We found no source for any successor product: LawSites reported on December 11, 2020 that the ROSS platform would no longer be available after January 31, 2021 and named no successor or acquirer. The phrase was removed. (1 page)

Source

Third Circuit opinion, Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153, filed September 29, 2026 (read October 30, 2026); LawSites, ROSS to shut down under pressure of Thomson Reuters lawsuit, December 11, 2020 (read October 30, 2026)

Pages that carry this correction (2)

Mata v. Avianca, Park v. Kim and Gauthier v. Goodyear: details corrected against the courts' own orders

· 12 pages

What changed

  • Mata v. Avianca: the June 22, 2023 order sanctions three respondents, attorneys Peter LoDuca and Steven Schwartz and their firm Levidow, Levidow & Oberman, with one $5,000 penalty, jointly and severally. Pages said two attorneys, both attorneys, one attorney or a supervising partner. We also removed a quotation from Schwartz's affidavit that is not in the order and now describe the screenshots the order does describe. (4 pages)
  • Removed 'disciplinary referral' from descriptions of Mata v. Avianca (the order makes none) and 'binding' or 'controlling' 2nd Circuit precedent from the New York, New Mexico and Georgia district pages: Mata is a Southern District of New York decision and does not bind other district courts. (8 pages)
  • Park v. Kim: removed 'potential referral to the admission committee' and 'translated into Korean if needed', which are not in the opinion. The Second Circuit referred Attorney Lee to its Grievance Panel and ordered her to give her client a copy of the decision. (1 page)
  • Gauthier v. Goodyear: the docket number is 1:23-CV-281, not 1:23-cv-396. The order sanctions the attorney ($2,000, one CLE hour, a copy to the plaintiff) for filing without reading the cases or confirming they existed; it does not hold that checking AI with a second AI tool is insufficient. That reading is ours and the page now says so. (1 page)

Source

Mata v. Avianca, S.D.N.Y. No. 22-cv-1461, Opinion and Order on Sanctions, ECF 54, June 22, 2023 (read October 30, 2026); Park v. Kim, 2d Cir. No. 22-2057, per curiam opinion, January 30, 2024 (read October 30, 2026); Gauthier v. Goodyear, E.D. Tex. No. 1:23-CV-281, Memorandum and Order, ECF 48, November 25, 2024 (read October 30, 2026)

Pages that carry this correction (12)

Lexis+ with Protégé trial: the Protégé review, the top-10 comparison and the Spellbook vs Protégé page no longer say the trial terms are not described

· 3 pages

What changed

  • Three pages said LexisNexis only offers a "Request a trial" button for Protégé, with terms not described, while our Lexis+ free-trial guide and other pages said 2 days. LexisNexis's own sign-up page is titled and headed as a 2-day free trial of Lexis+ with Protégé, and its product page says "Get your 2-day free trial" next to the "Request a trial" button. The three pages now say 2-day free trial, requested through a form. We did not find card or eligibility rules in the text we read, and the pages say so. (3 pages)

Source

LexisNexis, Lexis+ with Protégé free trial sign-up (read October 30, 2026); LexisNexis, Lexis+ with Protégé product page (read October 30, 2026)

Pages that carry this correction (3)

AI for healthcare law: HIPAA penalty amounts updated to the January 2026 HHS adjustment, and the Security Rule change described as a proposal

· 1 page

What changed

  • The page said HIPAA violations carry penalties up to $50,000 per incident with annual maximums of $2.1 million, and that a 2025 HIPAA Security Rule update added AI-specific risk assessment requirements. It now gives the amounts in HHS's January 28, 2026 inflation adjustment: $145 to $73,011 per violation in the first three tiers, a $73,011 minimum for uncorrected willful neglect, and a $2,190,294 calendar-year cap. The Security Rule change is a proposal (90 FR 898), its text does not use the phrase 'artificial intelligence', and HHS's agenda lists final action as 07/00/2027, so the claim of an AI-specific requirement in force was removed, together with an unsourced statement that OCR had signaled heightened scrutiny of AI violations. (1 page)

Source

HHS, Annual Civil Monetary Penalties Inflation Adjustment, Federal Register, January 28, 2026; HHS, HIPAA Security Rule To Strengthen the Cybersecurity of Electronic Protected Health Information, proposed rule, 90 FR 898 (January 6, 2025); reginfo.gov, RIN 0945-AA22 (final action listed as 07/00/2027)

Pages that carry this correction (1)

Harvey contract terms: pages no longer state annual contracts, seat minimums, a 25-seat minimum or a paid 60-90 day pilot as fact, because Harvey publishes none of them

· 27 pages

What changed

  • Replaced 'Harvey sells on annual enterprise contracts with seat minimums', 'Harvey's 25-seat minimum', 'some firms report a paid Harvey pilot of 60-90 days' and 'lawyers started using it within 60-90 days of contract signing' with what Harvey's own site shows: no price, seat minimum, contract length or pilot terms are published, and its Platform Agreement only says fees and the term end date are set in the order form. The pricing-page '403' remark was also fixed: harvey.ai/pricing returns 404 and we found no pricing page. Third-party estimates that stay are labeled unconfirmed. We did not contact Harvey. (27 pages)

Source

Harvey, Platform Agreement (last updated January 9, 2026; read October 30, 2026), where fees and the Term End Date are set in each customer's order form; Harvey, Request a demo (read October 30, 2026), a sales form with no seat minimum, term or pilot terms

Pages that carry this correction (27)

CoCounsel pricing page title no longer calls the whole product quote-only

· 1 page

What changed

  • The title, social titles and structured data of the CoCounsel pricing page said CoCounsel has no list price and is quote-only, although the body already gave Thomson Reuters' published in-house prices. Thomson Reuters publishes three in-house plans with starting prices; only law-firm plans, which show no number until a configurator is filled in, are quoted. The title now says that. (1 page)

Source

Thomson Reuters, CoCounsel plans for in-house legal teams: Essentials from $340.85, Practical Law Dynamic Tool Set with CoCounsel Essentials from $882.30, CoCounsel Legal from $1,694.90 per month (read October 30, 2026)

Pages that carry this correction (1)

Vincent AI, Harvey security, Lexis+ Protégé, Westlaw, Copilot and BigLaw pages re-read against the vendors' own pages

· 7 pages

What changed

  • On the vLex Vincent AI page, the vLex site calls the product Vincent by Clio, and its page titles read vLex (Part of Clio). The page said the site labels it Vincent by vLex (part of Clio). (1 page)
  • Harvey security page: removed TLS 1.2+/AES-256 and Azure FedRAMP details that Harvey's security page does not state, and the Claude Team and ChatGPT Team checklists (no SOC 2, no data residency, SOC 2 Type II, limited residency) that no vendor page supports. Claude Pro model training is opt-out on Anthropic's pricing page, not a commitment. (1 page)
  • Harvey vs Lexis+ Protégé: LexisNexis says Protégé pricing is customized (not a published price) and that teams can create, share and deploy AI agent skills, and its February 24, 2026 release mentions custom workflow design, so the page no longer says Protégé cannot build custom agents. Removed the 300+ workflows and 700,000+ daily agent tasks figures, which neither vendor page shows. (1 page)
  • Harvey vs Westlaw AI (CoCounsel): removed 'prices out 90% of the market', the claim that Harvey prioritizes Am Law 200 firms and bars small firms, and 'add-on to any Westlaw subscriber, no application', none of which a primary source supports. Thomson Reuters lists CoCounsel for in-house teams from $340.85 per month with a free trial. (1 page)
  • What AI tools BigLaw firms use: Harvey says 80% of Am Law 100 firms use it (replaces 'over 40'); Lexis+ with Protégé reached general availability on February 24, 2026 (not 2024); removed a Luminance document count, a Harvey seat minimum, a Copilot price Microsoft does not list and a claim that Am Law 200 firms are 2-3 years behind. (2 pages)
  • Harvey for small firms: Anthropic lists a context window of up to 1M tokens depending on the model, not 200K. Removed the 3-6 month sales cycle, the 7-12% of revenue figure, the 2,000-3,000 organizations count and the 70-80% claim, none of which has a source. (1 page)
  • Claude vs Copilot for lawyers: Microsoft lists the Copilot Business add-on at $18 per user per month billed yearly (promotion through 2026-12-31, $21 list), not the figure the page gave, so the combined-cost figures were recalculated. Removed the 90.9% BigLaw Bench score, a 128K Copilot context window, a GCC/FedRAMP comparison and Claude Enterprise zero-retention claims, which no vendor page states. Microsoft lists Harvey as a partner-built plugin for Copilot; it does not describe a training-data partnership. (1 page)

Source

vLex, Vincent AI and Fastcase Library pages (read October 30, 2026); Harvey, Security (read October 30, 2026); LexisNexis, Protégé product page (read October 30, 2026); LexisNexis pressroom, General Availability of Lexis+ with Protégé, February 24, 2026 (read October 30, 2026); Microsoft, Microsoft 365 Copilot plans and pricing (read October 30, 2026); Microsoft, Introducing Microsoft Copilot for legal, October 15, 2026 (read October 30, 2026); Anthropic, Claude pricing (read October 30, 2026); Thomson Reuters, CoCounsel plans for in-house teams (read October 30, 2026); Harvey, raise announcement, September 9, 2026 (read October 30, 2026)

Pages that carry this correction (7)

C.D. California page no longer says no judge there has an AI standing order

· 1 page

What changed

  • The page said no judge of the Central District of California had issued an AI standing order. The court's own judge pages show at least three that have: Judge Scarsi requires a separate declaration in civil cases, Judge Hsu requires one for AI-generated filing content, and Judge Hwang requires one in criminal cases and Rule 11 verification in civil cases. The page, its bottom line and two FAQ answers now say so. The statement that the district has no district-wide AI rule was re-checked and stands. (1 page)

Source

C.D. Cal. Judge Mark C. Scarsi, Initial Standing Order for Civil Cases (October 2026), paragraph 15, Use of Generative Artificial Intelligence (read October 30, 2026); C.D. Cal. Judge Anne Hwang, Standing Orders for Civil Cases (October 21, 2026) and Criminal Cases (September 25, 2026), paragraph 5 (read October 30, 2026); C.D. Cal. Judge Wesley L. Hsu, Standing Order for Newly Assigned Civil Cases, section K, Use of Artificial Intelligence (revised July 23, 2026) (read October 30, 2026)

Pages that carry this correction (1)

Judge Richard Andrews page no longer says he has an AI standing order

· 1 page

What changed

  • The title and description of the Judge Richard Andrews page presented an AI standing order from him, with disclosure, verification and sanctions terms. His page on the court's website lists only a 2017 order on responses to defense motions in criminal cases and a 2019 order letting parties use a word count in briefs, and neither mentions AI. The title, description and heading now say we found no AI standing order from him. (1 page)

Source

U.S. District Court for the District of Delaware, Judge Richard G. Andrews page and its Orders/Standing Orders list (read October 30, 2026); District of Delaware, general orders list, newest entry September 28, 2026 (read October 30, 2026)

Pages that carry this correction (1)

Westlaw pricing wording on 15 more pages: Thomson Reuters lists online Westlaw Advantage pricing for firms of up to 10 attorneys, so the pages no longer say it publishes no price or sells only by quote

· 15 pages

What changed

  • Replaced 'Thomson Reuters publishes no price', 'Westlaw is quote-only' and 'Westlaw is sold by quote' with what Thomson Reuters' own plans-pricing page shows: online Westlaw Advantage pricing for law firms of up to 10 attorneys, with larger firms sent to sales. Lexis and vLex still publish no price. No dollar figures were added. Wording about Westlaw Edge and the Westlaw Precision + CoCounsel bundle, which that page does not cover, is unchanged. (15 pages)

Source

Thomson Reuters, Westlaw Advantage plans and pricing for law firms (read October 29, 2026)

Pages that carry this correction (15)

Westlaw pricing wording, second sweep: a meta description, a source list, an FAQ answer and the Westlaw Precision paragraphs no longer say Thomson Reuters publishes no price or sells only by quote

· 7 pages

What changed

  • Replaced 'Westlaw Advantage is quote-only', 'Westlaw and Lexis sell by quote', 'Westlaw Precision is quote-only' and 'Thomson Reuters doesn't publish Westlaw Precision AI pricing' (in a search description, an FAQ answer, a pricing section and a vendor-comparison list) with what Thomson Reuters' own plans-pricing page shows: online Westlaw Advantage pricing for law firms of up to 10 attorneys, with larger firms sent to sales. Lexis and vLex still publish no price, and we found no published price for Westlaw Precision. No dollar figures were added. Wording about Westlaw Edge and the Westlaw Precision + CoCounsel bundle is unchanged. (7 pages)

Source

Thomson Reuters, Westlaw Advantage plans and pricing for law firms (read October 29, 2026)

Pages that carry this correction (7)

Westlaw is no longer described as publishing no price: Thomson Reuters lists online Westlaw Advantage pricing for firms of up to 10 attorneys

· 3 pages

What changed

  • Replaced 'Westlaw is sold by quote' and 'neither Westlaw nor Lexis publishes a price' with what Thomson Reuters' own plans-pricing page shows: online Westlaw Advantage pricing for law firms of up to 10 attorneys, with larger firms sent to sales. Lexis and vLex still publish no price. (3 pages)

Source

Thomson Reuters, Westlaw Advantage plans and pricing for law firms (read October 29, 2026)

Pages that carry this correction (3)

Western District of Oklahoma AI disclosure page rewritten from the text of Chief Judge Palk's order; unsourced 'mirrors the Northern District of Texas model' removed

· 4 pages

What changed

  • Pages said Judge Palk's order mirrors or is modeled on the Northern District of Texas (Judge Starr) order. The order we read does not mention that model, so the claim is gone. The Western District page now describes only what the order says: any party, pro se or represented, who uses generative AI to prepare a filed document must disclose that AI was used and the specific tool, and certify that they checked accuracy, including all citations and legal authority; failure may result in sanctions. (4 pages)
  • The Western District page also carried a garbled statistic from the NYC Bar survey, a claim that the district had no AI sanctions case, and a comparison with Kansas, Colorado and the Northern District that we had not checked. All were removed; the page now says we did not search WDOK dockets and did not verify other WDOK judges' orders. (1 page)

Source

Chambers of Chief U.S. District Judge Scott L. Palk, Disclosure and Certification Requirements (Generative Artificial Intelligence), W.D. Okla. website PDF, read 2026-10-29

Pages that carry this correction (4)

Published and Sources checked dates on three comparison pages corrected from October 21 to October 28, 2026

· 3 pages

What changed

  • These three pages said they were published and that their sources were checked on October 21, 2026. They were first published on October 28, 2026, and the vendor pages they cite were read that day. The visible dates and the Article dates in the page data now say October 28, 2026. (3 pages)

Source

Vendor pages read for these comparisons, October 28, 2026 (see each page's Sources list)

Pages that carry this correction (3)

Unsourced monthly cost ceiling removed from the Eve vs DIY, Harvey vs Westlaw AI and legal conferences pages

· 3 pages

What changed

  • Removed an unsourced monthly cost ceiling (500 dollars a month) from three places: the total cost of a DIY contract review workflow, a per-user budget cutoff for choosing between two vendors, and an example question in the conference-ROI advice. None has a public source: AI providers publish per-token rates, not a monthly total for a firm, and we have not benchmarked one. The pages now say the cost is unmeasured. (3 pages)

Source

Anthropic API pricing (per-token rates, no monthly total for a firm), read 2026-10-29

Pages that carry this correction (3)

Unsourced 500-dollar monthly cost ceiling removed from the Harvey alternatives page

· 1 page

What changed

  • Removed an unsourced claim that a 10-attorney firm could build a functional equivalent of Harvey for under 500 dollars a month with API costs. Anthropic publishes per-token API rates, not a monthly total for a firm, and we have not benchmarked one. The page now says the cost scales with usage and that any such figure is unverified. (1 page)

Source

Anthropic API pricing (per-token rates, no monthly total for a firm), read 2026-10-29

Pages that carry this correction (1)

AI-Generated Contracts: UETA section 14 now cited from the Uniform Law Commission text

· 1 page

What changed

  • The three UETA section 14 quotations were taken from Minnesota's enactment. They now cite the Uniform Law Commission's 1999 Final Act, which carries the same wording; Minnesota's statute is named as one state's version. (1 page)
  • Corrected a quotation attributed to the UETA. The Act's comments say a computer may be able to “learn through experience, modify the instructions in their own programs, and even devise new instructions”, quoting Allen and Widdison (1996); the page now uses those words. (1 page)

Source

Uniform Electronic Transactions Act (1999), Final Act, Uniform Law Commission

Pages that carry this correction (1)

Lexis pricing, EvenUp, contract-review ranking and Andersen v. Stability AI: five unsupported or misattributed figures and wordings fixed after re-reading their sources

· 4 pages

What changed

  • The Lexis pricing page presented a per-user monthly price range as a third-party estimate by The Legal Prompts. That guide only lists the range as what another site quotes and says it found no per-user price from LexisNexis, so we removed the range and say we have no verified per-seat price. (1 page)
  • The contract-review ranking attributed an Ironclad median of about $40,000 over 363 purchases to a procurement site that does not carry it. Vendr's own page now shows a median of $40,110 a year over 371 purchases, so the page cites that. (1 page)
  • The ranking gave Kira a median and a deal count that belong to Ironclad on CostBench; Kira's own row rests on only 3 deals, too few to quote, so the page now says that and drops the figure. (1 page)
  • The EvenUp profile said the tool cuts paralegal work on chronologies and demand letters from 20 to 40 hours to a few hours. No source supports those hours, so the page now calls the time saving the vendor's claim, with no independent measurement. (1 page)
  • The EvenUp profile said a draft demand letter arrives in 24 to 48 hours versus 2 to 4 weeks by hand. We found no source for either figure and removed them. (1 page)
  • The EvenUp profile gave case-volume thresholds (200+ cases, 10 versus 2 settlements a month) as the audience cut-off. No source supports them, so they were removed. (1 page)
  • The Andersen page said the court denied the defendants' motions to dismiss in August 2024. The August 12, 2024 order granted them in part and denied them in part: direct infringement, inducement and trademark claims survived and the DMCA Section 1202 claims were dismissed with prejudice. The page now says so and no longer restates reasoning we had not read in the order. (1 page)
  • The Andersen page described a DMCA Section 1202 theory as part of the case and placed Getty's suit in Delaware. The 1202 claims were dismissed in August 2024 and the Getty venue statement is no longer reliable, so both were rewritten. (1 page)

Source

The Legal Prompts, Lexis AI pricing (lists $125-275 only as what a third site quotes; no literal per-user price from LexisNexis), read 2026-10-29; Vendr, Ironclad pricing (median $40,110 per year, 371 purchases), read 2026-10-29; CostBench, Ironclad vs Kira Systems (Kira $35,000 median on 3 deals; Ironclad $38,825 on 328 deals), read 2026-10-29; Copyright Alliance, Top Takeaways from the Order in Andersen v. Stability AI (August 12, 2024 order granting in part and denying in part), read 2026-10-29

Pages that carry this correction (4)

Unsourced majority claims about state bars claims and 'at least 8 lawyers sanctioned' replaced by named opinions, a dated source or removed

· 17 pages

What changed

  • These pages said that 'a majority of state bars' permit AI use, require notification, approve AI CLE credit, agree on cannabis practice, have addressed chatbots, publish opinions by email, avoid fee reductions, give free Fastcase access, or are 'more detailed' than Oregon's opinion, with no source and no date. Each is now either a statement about the opinions we re-read on October 29, 2026 (Texas 705, North Carolina 2024 FEO 1, D.C. 388, Kentucky E-457: they tie AI use to competence and confidentiality duties; Kentucky finds no duty to disclose rote AI research, Texas and D.C. allow billing only for time actually spent), Fastcase's own statement of 80+ national, state and county bar partners, a pointer to the reader's own state rules, or our dated state-by-state index. The ABA and Florida Bar pages answered 403 that day and were not re-read; the Florida and ABA positions on informed consent are reported as Texas 705 describes them. (15 pages)
  • Two pages said 'at least 8 lawyers' had been sanctioned for AI-related misconduct, with no source. They now say lawyers have been sanctioned for AI-fabricated citations, beginning with Mata v. Avianca (S.D.N.Y., June 22, 2023, a $5,000 penalty), which we re-read. (2 pages)

Source

Texas Professional Ethics Committee Opinion 705, re-read 2026-10-29; North Carolina State Bar 2024 Formal Ethics Opinion 1, re-read 2026-10-29; District of Columbia Bar Ethics Opinion 388, re-read 2026-10-29; Kentucky Bar Association KBA E-457, re-read 2026-10-29; Fastcase: partners with over 80 national, state and county bar associations, read 2026-10-29; Mata v. Avianca, S.D.N.Y. Opinion and Order, June 22, 2023 ($5,000 penalty), re-read 2026-10-29; AI Vortex state-by-state AI ethics index, primaries re-read 2026-10-28

Pages that carry this correction (17)

Alaska, New Mexico, West Virginia and Arizona AI law pages: 'no ethics opinion / no guidance / silent' removed; each bar or court body has published AI guidance

· 4 pages

What changed

  • The Alaska AI law page said the Bar had issued no ethics opinion and no bar guidance and called the state's posture silent. The Alaska Bar Association's Ethics Opinion 2025-1 on generative AI was approved on April 3, 2025 and adopted on April 23, 2025. We read it on October 29, 2026; the page now describes it. We did not verify any Alaska task force, legislation or court rule. (1 page)
  • The New Mexico AI law page said the State Bar had issued no ethics opinion, informal guidance or publication on AI. The Ethics Advisory Committee issued Formal Opinion 2024-004 on September 24, 2024. We read it on October 29, 2026; the page now describes it. We did not verify any New Mexico court rule on AI. (1 page)
  • The West Virginia AI law page said no formal ethics opinion had followed and that guidance was informal. The Lawyer Disciplinary Board approved Legal Ethics Opinion 24-01, Artificial Intelligence, on June 14, 2024, and the Office of Disciplinary Counsel lists it. We read it on October 29, 2026; the page now describes it. (1 page)
  • The Arizona AI law page said the State Bar had issued no formal AI guidance and called the state silent. The State Bar of Arizona published practical guidance on generative AI (undated and unnumbered) and the Arizona Supreme Court's AI steering committee issued best practices for lawyers and judges on November 14, 2024. We found no numbered ethics opinion, which the page now says. We read the State Bar guidance on October 29, 2026. (1 page)

Source

Alaska Bar Association, Ethics Opinion 2025-1, Generative Artificial Intelligence & The Practice of Law (approved Apr. 3, 2025; adopted by the Board of Governors Apr. 23, 2025; read 2026-10-29); State Bar of New Mexico Ethics Advisory Committee, Formal Opinion 2024-004, Using Generative Artificial Intelligence in the Practice of Law (issued Sept. 24, 2024; read 2026-10-29); West Virginia Office of Disciplinary Counsel, Legal Ethics Opinions page, L.E.O. 24-01 Artificial Intelligence (read 2026-10-29); West Virginia Lawyer Disciplinary Board, Legal Ethics Opinion 24-01, Artificial Intelligence (approved and entered June 14, 2024; scanned PDF read 2026-10-29); State Bar of Arizona, Guidance for the Use of Generative Artificial Intelligence in the Practice of Law in Arizona (undated; read 2026-10-29); Arizona Supreme Court Steering Committee on Artificial Intelligence and the Courts, Generative AI: Ethical Best Practices for Lawyers and Judges (Nov. 14, 2024; read 2026-10-29)

Pages that carry this correction (4)

Rhode Island AI law page: 'no formal AI guidance' removed; the Supreme Court amended Rule 1.1 Comment [6] and adopted advisory Interim Generative AI Guidelines on June 17, 2026

· 1 page

What changed

  • The Rhode Island AI law page said the state had no formal AI guidance and was silent. On June 17, 2026 the Rhode Island Supreme Court amended Comment [6] to Article V, Rule 1.1 and adopted advisory Interim Guidelines for the Ethical Use of Generative AI by Rhode Island Lawyers. The Court's order says the Guidelines do not amend the Rules of Professional Conduct and do not carry the force of law. We read the order on October 29, 2026; the page now describes it. We did not verify whether the Rhode Island Bar Association has issued its own AI materials, or any Rhode Island legislation, task force or court rule on AI disclosure. (1 page)

Source

Rhode Island Supreme Court, Order amending Article V, Rule 1.1 and adopting Interim Generative AI Guidelines for lawyers and judicial officers (June 17, 2026; read 2026-10-29)

Pages that carry this correction (1)

Ohio AI law page: 'no formal AI guidance / silent' removed; the Board of Professional Conduct published the Ohio Ethics Guide on AI

· 1 page

What changed

  • The Ohio AI law page said Ohio had no formal AI guidance and that the state bar was silent. The Ohio Board of Professional Conduct has published the Ohio Ethics Guide: Artificial Intelligence for Lawyers and Judicial Officers. It is nonbinding staff guidance, not a Board opinion, and carries no opinion number. We read it on October 29, 2026; the page now describes it. We did not verify any Ohio legislation, task force or court rule on AI, and we did not review the Supreme Court of Ohio's AI Resource Library that the guide cites. (1 page)

Source

Ohio Board of Professional Conduct, Ohio Ethics Guide: Artificial Intelligence for Lawyers and Judicial Officers (nonbinding staff guidance, 2026; read 2026-10-29)

Pages that carry this correction (1)

Best AI for legal drafting: Spellbook per-user estimate removed; Claude vs ChatGPT for legal: Gemini 3.8 Flash dates now confirmed on Google's pages

· 2 pages

What changed

  • The drafting comparison quoted a third-party per-user monthly price range for Spellbook, attributed to a Bind Legal page. That page no longer carries the figure and says Spellbook does not publish rates, so we removed the estimate and say we have no verified per-user price. (1 page)
  • The Claude vs ChatGPT comparison said Gemini 3.8 Flash's general availability (2026-09-02) and introductory rate (through 2026-12-31) were reported by third parties and not confirmed on Google's pages. Google's changelog and pricing page now confirm both, so the sentence cites them. (1 page)

Source

Bind Legal, Spellbook pricing 2026 ('Spellbook does not publish rates... routes you to a free trial or a demo'), read 2026-10-29; Google Gemini API changelog (September 2, 2026: Gemini 3.8 Flash generally available), read 2026-10-29; Google Gemini API pricing (Gemini 3.8 Flash $0.75 input and $3.75 output per million tokens through December 31, 2026; $1.50 and $7.50 from January 1, 2027), read 2026-10-29

Pages that carry this correction (2)

Unsourced claims reworded on the Claude Projects, Harvey Agent Builder and Judge Starr pages

· 3 pages

What changed

  • Removed a fixed 200K-token, 150-page limit for Claude project knowledge. Anthropic's help page describes retrieval (RAG) mode that expands capacity up to 10x and gives no fixed page count, so the page now says that. (1 page)
  • Removed an unsourced 80%-of-capability-for-10%-of-cost claim and a vendor price range, and replaced a statement that team members see each other's conversations: Anthropic's help page does not say that. (1 page)
  • Harvey Agent Builder guide: removed an unsourced 80/20 split, a claim about Harvey's documentation and a claim about seat minimums; the page now cites only what Harvey published (human-in-the-loop checkpoints, no public price). (1 page)
  • Judge Starr page: removed two unsourced statements (that dozens of judges followed the 2023 order, and who the disclosure covers) and added that under LR 7.2(f) the presiding judge may direct the filer to identify the parts prepared with generative AI. (1 page)

Source

N.D. Tex. Local Civil Rules, LR 7.2(f), effective September 1, 2026, read 2026-10-29; Anthropic Help Center: What are projects? (RAG mode), read 2026-10-29; Harvey: Introducing Agent Builder (March 9, 2026), read 2026-10-29

Pages that carry this correction (3)

North Carolina page: unsourced 'most detailed', 'binding', 'disciplinary' and 'hallucination' claims removed; cannabis page now cites California's opinion from the State Bar's own page

· 2 pages

What changed

  • The North Carolina page called 2024 FEO 1 one of the most detailed (and most specific) state bar AI opinions in the country, said it creates binding obligations with disciplinary consequences, and said it requires attorneys to understand hallucination risks. We re-read the opinion: it makes no comparison with other states, does not discuss discipline, and does not use the word hallucination (it says AI work product can be inaccurate or unreliable despite its appearance of reliability). The page now describes what the opinion says and attributes the duties to the Rules it interprets. (1 page)
  • The page said submitting fabricated citations violates Rule 3.3 and that putting client data in an unvetted AI platform violates Rule 1.6. The opinion cites Rule 3.1 and N.C. Rule of Civil Procedure 11 for pleadings and asks for reasonable efforts under Rules 1.6(c) and 5.3; the page now says that. (1 page)
  • The cannabis page described California's opinion through a secondary source and said state conclusions are not uniform. We read the State Bar of California's opinion on calbar.ca.gov: it says a lawyer may advise on and assist conduct permitted by California cannabis law despite federal law, may not help avoid detection, and notes similar approaches in virtually every jurisdiction that has legalized cannabis, with differing details. The page now says that. (1 page)

Source

North Carolina State Bar 2024 Formal Ethics Opinion 1 (adopted November 1, 2024), re-read 2026-10-29; State Bar of California Formal Opinion on advising cannabis clients, calbar.ca.gov, read 2026-10-29

Pages that carry this correction (2)

Legora reviews and comparison: unsourced 'built on Claude only', 'chose Claude over GPT' and 'Harvey is GPT-based' removed; the pages now say which models vendors publish

· 3 pages

What changed

  • Three pages said Legora is built only on Anthropic's Claude, that it chose Claude over GPT, and that Harvey is built on GPT. We found no vendor source for any of these. Microsoft's customer story of May 27, 2025 says Legora uses multiple models, including Azure OpenAI's GPT-4o and o1; Anthropic's page quotes Legora's CTO on building with Claude; Legora's security page lists no models. The pages now say that the sources point to more than one provider and suggest asking the vendor in writing which providers process client data. (3 pages)

Source

Microsoft, Legora customer story on Azure OpenAI Service (May 27, 2025), says Legora uses multiple models, including Azure OpenAI's GPT-4o and o1; Anthropic, Claude for Legal (read 2026-10-29), quotes Legora's CTO on building with Claude, with no statement that Claude is the only model; Legora, security page (read 2026-10-29), lists no models or model providers

Pages that carry this correction (3)

Judge Altman's page now places his chambers in Miami and reports that chambers has no standing order

· 1 page

What changed

  • Replaced 'his courtroom in Fort Lauderdale' with the address on his chambers page (Wilkie D. Ferguson Jr. Courthouse, Miami) and added that the page states chambers does not have a standing order. (1 page)

Source

Judge Roy K. Altman, chambers page, S.D. Fla. (read October 29, 2026)

Pages that carry this correction (1)

Johnson v. Dunn: described as an AI-evidence case; the order is about ChatGPT-fabricated citations and the disqualification of three Butler Snow lawyers

· 1 page

What changed

  • The title, description, headline and two FAQ answers said the case raised AI-generated evidence admissibility and an expert report under FRE 901. That was wrong: no AI evidence was offered. The order sanctions lawyers for fabricated citations that ChatGPT produced for two motions. The title, description, headline and FAQ now say what the court decided on July 23, 2025. (1 page)
  • The page gave the wrong docket number (2:23-cv-01492; the order is 2:21-cv-1701-AMM), said the clerk notified bar regulators in every state where the lawyers are licensed (the order directs service on the Alabama State Bar and other applicable licensing authorities), and misquoted the court ("recklessness in the extreme" about Reeves). All three were corrected against the order. (1 page)

Source

Johnson v. Dunn, No. 2:21-cv-1701-AMM, 792 F. Supp. 3d 1241 (N.D. Ala. July 23, 2025), sanctions order (read October 29, 2026)

Pages that carry this correction (1)

Harvey vs Copilot and Harvey vs Claude vs ChatGPT: Harvey's 'Security Addendum' is no longer called a 'data processing addendum'; the 25,000+ workflows are credited to Workflow Builder

· 2 pages

What changed

  • Both pages said Harvey's security page describes a 'data processing addendum'. The page names a 'Security Addendum' with binding terms on data protection, data access and incident response SLAs; we now use Harvey's own name. (2 pages)
  • The Copilot comparison attached Harvey's 25,000+ custom agentic workflows to Agent Builder. Harvey's March 9, 2026 post says they were built with Workflow Builder, which Agent Builder succeeds; the sentence now says so. (1 page)

Source

Harvey security page ('Harvey's Security Addendum includes binding terms on data protection, data access, incident response SLAs'), read 2026-10-29; Harvey, Introducing Agent Builder (Mar 9, 2026): 'built over 25,000 agentic custom workflows with Workflow Builder'; 'evolving Workflow Builder into Agent Builder', re-read 2026-10-29

Pages that carry this correction (2)

Harvey vs. Build-Your-Own Claude: firm size of Harvey's mid-sized customer examples corrected

· 1 page

What changed

  • Removed the statement that Harvey's three mid-sized customer examples were all under 200 lawyers. Harvey's Foley Hoag story says the firm has more than 310 lawyers; the page now gives each firm's size in Harvey's wording (nearly 100, more than 150 and more than 310) and says the figures are self-reported. (1 page)

Source

Harvey, Foley Hoag customer story (read October 29, 2026); Harvey, One Problem, Different Paths: How Mid-Sized Firms Build AI Adoption (October 21, 2026, read October 29, 2026)

Pages that carry this correction (1)

Luminance funding and Harvey fundraising totals tied to the companies' own announcements

· 2 pages

What changed

  • Replaced the unsourced $100M+ figure for Luminance with the round Luminance announced itself: a $75 million Series C led by Point72 on February 27, 2025, bringing total funding to over $115 million, per its own blog. (1 page)
  • Replaced 'Harvey has raised more than $1 billion (as of March 2026)', a total we could not trace to a primary source, with Harvey's two dated announcements: $200M at $11 billion (March 25, 2026) and $550M at $15.5 billion (September 9, 2026). (1 page)

Source

Luminance, Series C blog post (February 27, 2025); Harvey, $11B growth round announcement (March 25, 2026); Harvey, $550M round announcement (September 9, 2026)

Pages that carry this correction (2)

Anthropic, Harvey and Legora valuation claims tied to their own announcements; unsourced motives removed

· 3 pages

What changed

  • Anthropic's $965 billion post-money valuation is now dated (May 28, 2026) and attributed to its announcement. We removed the stated purpose of the round ('to build the best general-purpose AI model' / 'to build Claude'), which Anthropic's announcement does not state. (2 pages)
  • Removed the claim that Harvey's $15.5 billion valuation is what makes it the default for American practices; Harvey's September 9, 2026 announcement states the round and valuation only, and the page now cites exactly that. (1 page)
  • Removed our reading that Legora's $8.8 billion valuation reflects a market US-built tools underserve; Legora's announcement and the press coverage give the round and the company's own revenue figure, not that reason. The ARR line now says it is the company's own figure as TechCrunch reports it, instead of saying TechCrunch did not verify it. (1 page)

Source

Anthropic, Series H announcement (May 28, 2026); Harvey, $550M round announcement (September 9, 2026); Legora, $550M Series E announcement (October 21, 2026)

Pages that carry this correction (3)

Harvey valuation inferences and Relativity's valuation corrected or attributed against primaries

· 3 pages

What changed

  • Removed the claim that Harvey's $15.5 billion valuation is built on large-firm contracts: Harvey's September 9, 2026 announcement states the round and valuation only. The page now cites what the announcement does say. (1 page)
  • Removed our inference that Harvey has no trial because it is valued at $15.5 billion and does not need attorneys to validate its product; nothing in Harvey's announcement supports it. (1 page)
  • Relativity's $3.6 billion is a valuation reported in March 2021 (LawNext, citing the Wall Street Journal), not a current figure; the page now says so. (1 page)

Source

Harvey, $550M round announcement (September 9, 2026); LawNext, Investment in Relativity puts the e-discovery company's valuation at a reported $3.6 billion (March 19, 2021)

Pages that carry this correction (3)

Fastcase and vLex through bar associations: removed 'most bar memberships', 'dozens of states', '80%+ of research needs', 'not AI-powered', a standalone Fastcase subscription and 'basic vLex access'; the pages now use vLex's own wording

· 12 pages

What changed

  • The solo-practitioner guide said Fastcase or vLex through the bar 'covers 80%+ of research needs', that 'dozens of states' provide it and that most solos find it enough. We found no source for any of the three; the page now gives vLex's own count of partner bars and says to test it on your own matters. (1 page)
  • Pages said Fastcase is 'free with most bar memberships', 'not AI-powered' or has 'basic AI features'. vLex says it is available only through its 80+ partner bars and describes AI-powered case analysis in it; the pages now say that, with the read date. (3 pages)
  • Two pages priced a 'Fastcase subscription'. vLex's Fastcase Library page says it is available exclusively through bar partnerships and shows no subscription price, so the pages now say that and point to a written vLex quote for its other products. (2 pages)
  • The vLex Vincent AI page said bar members get 'basic vLex access' or the 'core vLex platform' at no cost and that the Fastcase merger added 'millions' of documents. vLex does not say the bar benefit includes Vincent or the rest of its platform, and we found no document count for the merger; the page now says what vLex says. (1 page)
  • Comparison pages said 'many state bars' or 'multiple bar associations' include vLex or Fastcase for free. The free benefit is Fastcase Library through the 80+ partner bars vLex lists, not vLex as a whole; the pages now say so. (5 pages)

Source

vLex, Fastcase Library ('available exclusively through bar association partnerships, serving over 1 million lawyers nationwide through 80+ state, county, and specialty bar associations'; 'complimentary access as part of your membership benefits'; AI-powered case analysis), read 2026-10-29; Fastcase ('Fastcase is now part of Clio'; 'free member benefit' through 80+ bar associations), read 2026-10-29

Pages that carry this correction (12)

eImmigration pricing page now lists prices: the 'no published price' statement replaced with its three listed plans

· 1 page

What changed

  • The page said eImmigration publishes no price and sends buyers to a demo request, which is what we read on October 20. Its pricing page now lists three plans per user per month, so the page gives the listed prices, the plan differences and the date read, and no longer calls the price quote-only. (1 page)

Source

eImmigration, pricing page (Starter $60 / $55, Essentials $90 / $82, Complete $110 / $100 per user per month, monthly / billed annually; read 2026-10-29)

Pages that carry this correction (1)

Doe v. GitHub: the Ninth Circuit affirmed the DMCA dismissal on September 16, 2026; the page said the DMCA risk could return and described the case as two remaining claims

· 1 page

What changed

  • The page said the DMCA claim failed because the statute does not apply without more specific allegations, that the DMCA risk could return with better-pleaded allegations, and that the case proceeds on two claims after 20 of 22 were dismissed. On September 16, 2026 the Ninth Circuit affirmed the dismissal of the DMCA Section 1202(b) claim on the ground that Copilot creates new works rather than removing attribution from copies of existing code, and described the breach of contract claims as pending. The page now states that holding, drops the 20-of-22 count we could not confirm, and says the ruling does not decide whether Copilot's output infringes or whether training is lawful. (1 page)

Source

Ninth Circuit opinion, Doe v. GitHub, Inc., No. 24-7700 (Sept. 16, 2026), read 2026-10-29

Pages that carry this correction (1)

Conseal profile now cites the vendor's privacy policy and FAQ for its on-device claim

· 1 page

What changed

  • Replaced the answer that told readers to confirm the on-device claim in the privacy terms, which we had not read. We have now read Conseal's privacy policy and FAQ: both say documents are processed on the device and that the company does not receive or store them. The answer attributes this to the vendor, notes that we did not test it, and mentions the Live Detect option and the crash reports sent only on request. (1 page)

Source

Conseal privacy policy (read October 29, 2026); Conseal FAQ (read October 29, 2026)

Pages that carry this correction (1)

AI CLE competence requirements: Mississippi 'first state to mandate AI CLE' claim withdrawn, '40 states' count removed

· 1 page

What changed

  • Removed the claim that Mississippi is the first state to mandate AI CLE. Mississippi's CLE rule and the Commission's FAQ require 12 hours a year, with 1 hour of ethics, professionalism and related topics, and do not mention AI; the Supreme Court's orders amending the rules on September 23, 2025 and June 10, 2026 add no AI credit. The claim had no source. The guide now names New Jersey's technology credit (cycle ending December 31, 2027) as the one state rule read at the source that names AI, and drops the unsourced California and New York statements. (1 page)
  • Removed the '40 states have adopted the duty of technology competence' count, which had no primary source on the page, and stated North Carolina's requirement precisely: 24 hours every two years, including one technology hour. (1 page)

Source

Mississippi Rules and Regulations for Mandatory Continuing Legal Education (read October 29, 2026); Mississippi Commission on CLE, FAQ (read October 29, 2026); Mississippi Supreme Court order No. 89-R-99011-SCT, September 23, 2025 (read October 29, 2026); Mississippi Supreme Court order No. 89-R-99011-SCT, June 10, 2026 (read October 29, 2026); North Carolina State Bar, CLE for members (read October 29, 2026)

Pages that carry this correction (1)

Claude Design guides no longer cite a documentation page that does not exist, and EvenUp pricing wording corrected

· 5 pages

What changed

  • Removed the citation to an Anthropic documentation page that returns 404 and the statements attributed to it: that Claude Design supports loadable skills and a design-tokens flow with spacing scales. Anthropic's help center documents design systems (colors, typography, components, layout patterns) and does not describe a user-loadable skill feature, so the pages now cite the help article and say that a "skill" here is a reusable brief a team writes itself. (4 pages)
  • Removed the statement that EvenUp's pricing page links subscription terms. That page now redirects to the homepage, which shows a demo button and no price. (1 page)

Source

Claude Help Center, Set up your design system in Claude Design

Pages that carry this correction (5)

AI in International Arbitration: unsourced figures and institution statements replaced

· 1 page

What changed

  • Removed the account of a 2024 ICC construction arbitration (4.2 million documents, an estimated $2.8 million saved in review costs). We could not find the case, the institution or any study behind it. (1 page)
  • Removed the 85-92% accuracy range for multi-language relevance classification, which cited peer-reviewed studies we could not identify. (1 page)
  • Removed the claim that about 30% of commercial arbitration awards are public, which had no source. The section now cites LCIA Rules Article 30.3: the LCIA publishes no award without the consent of all parties and the tribunal. (1 page)
  • Removed the statement that Brattle Group and Compass Lexecon use proprietary AI models on ICC and ICSID award databases, which we could not source. (1 page)
  • Removed the claim that AI translation of arbitration documents exceeds 90% accuracy; we could not find a published rate. (1 page)
  • Replaced the statement that ICC Article 22, LCIA Article 30 and SIAC Rule 39 all impose confidentiality obligations on parties and arbitrators. LCIA Article 30 binds the parties and the tribunal; ICC Article 22(3) lets the tribunal order confidentiality measures at a party's request. SIAC is no longer cited; for a SIAC case, read its rules on siac.org.sg. (1 page)
  • Replaced the statement that disclosing AI use is becoming the norm with the CIArb Guideline's position (para. 7.1): disclosure may be required where AI use could affect the evidence or the outcome. The Guideline is soft law that applies only if the parties adopt it. (1 page)
  • Removed the figure of 30,000 contracts reviewed in hours instead of months, which had no source. (1 page)

Source

LCIA Arbitration Rules (2020), Article 30; ICC Arbitration Rules (2021), Article 22; CIArb Guideline on the Use of AI in Arbitration (2025, updated September 2025)

Pages that carry this correction (1)

AI in International Arbitration: unsourced ICC, LCIA and SIAC guidance claims removed

· 1 page

What changed

  • Removed the statement that the ICC released a 2024 guidance note on AI. In the ICC Note to Parties and Arbitral Tribunals (updated June 1, 2026) AI appears only in the data-protection section, as a reference to the EU AI Act; we found no ICC rule on using AI tools in a case. The section now points to the SVAMC Guidelines of April 30, 2024. (1 page)
  • Removed the statements about LCIA arbitrator guidelines and SIAC's position on AI, which we could not verify against the institutions' own pages. (1 page)
  • Replaced the FAQ answer on disclosing AI use, which cited ICC and LCIA guidance and tribunal procedural orders we could not source, with what the SVAMC Guidelines say: disclosure is not necessary as a general matter and is decided case by case. (1 page)
  • Replaced the second FAQ answer on AI disclosure rules, which attributed guidance to the ICC and SIAC, with the SVAMC position, which applies only when the parties or the tribunal adopt it. (1 page)

Source

SVAMC Guidelines on the Use of Artificial Intelligence in Arbitration, 1st edition (published April 30, 2024); ICC Note to Parties and Arbitral Tribunals on the Conduct of the Arbitration (updated June 1, 2026)

Pages that carry this correction (1)

Alabama AI law page and N.D. Alabama page: 'no formal ethics opinion' removed; the Alabama State Bar issued Formal Opinion 2026-01 on AI use

· 2 pages

What changed

  • The Alabama AI law page said the State Bar had issued no formal ethics opinion on AI and called its guidance informal; the Northern District of Alabama page said the same. The Bar's Formal Opinions page lists Formal Opinion 2026-01, 'Artificial Intelligence Use: Best Practices Under Existing Professional Conduct Rules', and we read its text on October 29, 2026. It adds no new duties; it applies the existing rules to AI and addresses verification, confidentiality and informed consent, billing (time saved is not billable, review time is, subscriptions are overhead unless agreed) and client communication. The pages now say so, and the status label moved from partial guidance to has AI regulation. The PDF prints no issue date, so the pages do not state one. (2 pages)

Source

Alabama State Bar, Formal Opinions page: Formal Opinion 2026-01, Artificial Intelligence Use: Best Practices Under Existing Professional Conduct Rules (read 2026-10-29); Alabama State Bar, Formal Opinion 2026-01 (PDF, read 2026-10-29; prints no issue date)

Pages that carry this correction (2)

Alabama and Massachusetts court-disclosure pages: 'no State Bar guidance', 'fabricated citations violate Rule 3.3' and 'the SJC found existing rules sufficient' removed; Alabama's Formal Opinion 2026-01 now described

· 2 pages

What changed

  • Both pages said that submitting fabricated citations violates Rule 3.3. Alabama's State Bar opinion cites ABA Formal Opinion 512 for the point that submitting AI-generated work without independent verification can violate, at a minimum, Rules 1.1 and 3.3; Rule 3.3(a)(1) concerns knowingly false statements to a tribunal, so whether a filing breaches it depends on the facts. The pages now say that and cite court sanctions under Rule 11 where reported. (2 pages)
  • The Alabama page said the State Bar had issued no formal ethics opinion or guidance and that Alabama was waiting on a first AI sanctions case. The State Bar's Formal Opinion 2026-01 exists (the text states no adoption date and cites an April 24, 2026 Alabama Supreme Court decision, Ibach v. Stewart). The page now describes the opinion, says it does not itself create a filing-disclosure duty but tells lawyers to check each court's rules, and no longer claims that no court rule or standing order exists, because court rules are for each court to confirm. (1 page)
  • The Massachusetts page said the SJC Interim Guidelines apply to court staff only. Press and tracker summaries of the mass.gov text say they apply to judges, clerks, registers, recorders, employees, law clerks, interns and contractors of the state courts, are dated November 12, 2025 and are interim. The page now says that and tells readers to confirm on the SJC site. (1 page)
  • The Massachusetts page said the SJC concluded, and reasoned about procedural complexity, that existing professional conduct rules are sufficient for attorneys. As reported from Chief Justice Budd's December 9, 2025 address, the finding was the Standing Advisory Committee on the Rules of Professional Conduct's recommendation of no changes for now, not an SJC holding, and nothing we read gives the reasoning the page described. The page now attributes it that way. (1 page)

Source

Alabama State Bar, Formal Opinion 2026-01, Artificial Intelligence Use: Best Practices Under Existing Professional Conduct Rules (alabar.org), read 2026-10-29; Massachusetts SJC Interim Guidelines for Use of Generative AI (described from secondary reports, 2026-10-29)

Pages that carry this correction (2)

AI in Law Statistics 2026: ABA 2024 survey figure and firm-size tier corrected

· 1 page

What changed

  • Removed the 30.2% figure for AI use in the ABA 2024 Legal Technology Survey Report and the Lawyers Mutual link cited for it: that article does not contain the figure. The page now says about 30% of respondents use AI-based tools (up from 11% in 2023), as reported by the ABA Journal and LawNext, and gives the top firm-size tier as firms with 100 or more attorneys (the page said more than 500). It also states that these are secondary reports on the survey, not the report itself. (1 page)

Source

ABA Journal, ABA tech report finds AI adoption is growing (read October 29, 2026); LawNext, ABA Tech Survey finds growing adoption of AI in legal practice (March 7, 2025, read October 29, 2026)

Pages that carry this correction (1)

Unsourced 'ABA Ethics Opinion 20-480' cannabis cite removed; 'North Carolina's Unique Emphasis' on vendor vetting replaced by what 2024 FEO 1 says

· 2 pages

What changed

  • The cannabis law page said an 'ABA Ethics Opinion 20-480' confirmed that lawyers can advise cannabis clients in states where it is legal. We could not find that opinion in the sources we could read, and no other source we read cites it, so the citation is gone. The answer now says each state bar addresses the question itself, names the State Bar of California's May 2020 formal opinion as an example, and tells readers to check their own bar. (1 page)
  • The North Carolina page called vendor vetting 'North Carolina's Unique Emphasis' and said it sets the state apart. 2024 FEO 1, which we re-read, applies Rules 5.3 and 1.6(c) and the State Bar's earlier 2011 opinion on third-party technology to AI vendors the way it would to any third-party software, and does not claim to be different from other state bars. The heading is now 'Vendor Vetting Under the Opinion' and the section describes what the opinion says. (1 page)

Source

North Carolina State Bar 2024 Formal Ethics Opinion 1 (adopted November 1, 2024), re-read 2026-10-29; State Bar of California May 2020 formal opinion on advising cannabis clients, as described by the California Lawyers Association (July 1, 2020), read 2026-10-29

Pages that carry this correction (2)

xAI v. OpenAI page metas no longer say the suit was filed in August 2025 or cite '8 former xAI employees'; vLex bar count on two practice pages now uses vLex's own figure

· 3 pages

What changed

  • The search description, social previews and tags of the xAI v. OpenAI page said xAI filed the suit in August 2025 and cited 8 former employees as fact. The complaint was filed in September 2025 (OpenAI's motion to dismiss is dated October 2, 2025), and the count of engineers is Reuters' figure, so the metadata now says September 2025 and attributes the engineers to Reuters, matching the body. (1 page)
  • The AI for solo lawyers and AI for nonprofit legal pages said over 35 state bars offer vLex Vincent AI as a member benefit. We could not source that number. The pages now give vLex's own figure, 80+ state, county, and specialty bar associations for its Fastcase Library, attributed to vLex, and tell readers to check whether their bar's benefit includes Vincent AI. (2 pages)

Source

Reuters via Yahoo Finance, Judge dismisses xAI trade secrets lawsuit against OpenAI (February 25, 2026); OpenAI motion to dismiss, X.AI Corp. v. OpenAI, No. 3:25-cv-08133-RFL (N.D. Cal.), Dkt. 30 (October 2, 2025); vLex, Fastcase Library (read October 28, 2026)

Pages that carry this correction (3)

Whiting v. City of Athens: unverified quotations and an unverified case name removed

· 1 page

What changed

  • The page put several quotations in the court's mouth, named a fabricated case ("Berg v. Knox Cnty."), said the court applied Federal Rule of Appellate Procedure 46(c) and Rule 11, and called the opinion published and binding. We could not find any of this in the excerpt of the opinion reproduced by Reason/Volokh (March 14, 2026), and the order itself is not accessible to us. The quotations, the case name and the unsupported claims are removed; the page now describes only what that excerpt supports (Rule 38 and the court's inherent authority, a $15,000 fine on each attorney, double costs and fee reimbursement), and says that we have not confirmed whether the opinion is designated for publication. (1 page)

Source

Reason/Volokh, Lawyers citing nonexistent cases ordered to pay opponent's attorney fees, double costs, $15K fine (March 14, 2026)

Pages that carry this correction (1)

Whiting v. City of Athens: figures and case number checked against the opinion itself

· 1 page

What changed

  • We now read the opinion itself, not only the Reason/Volokh excerpt. The case number is Nos. 24-5918/5919/25-5424 (Justia files the PDF under 25-5425 in its URL). The court counts "over two dozen" fake citations and misrepresentations of fact and calls that a conservative estimate, so the earlier "24" is removed. The $15,000 fine on each attorney (paid to the court's registry), double costs and reimbursement of the appellees' attorneys' fees are confirmed. The case name Berg v. Knox Cnty. we removed earlier as unverified is in fact listed in the court's appendix as a citation it could not find, cited once (page 16 of the brief in No. 25-5424), not six times as an older version of the page said; it is back, attributed to the opinion. The opinion is marked "Recommended for publication". (1 page)

Source

Whiting v. City of Athens, Nos. 24-5918/5919/25-5424 (6th Cir. Mar. 13, 2026), opinion PDF on Justia

Pages that carry this correction (1)

Whiting v. City of Athens: sanction amount and citation count now attributed to Reason/Volokh; primary order not accessible

· 1 page

What changed

  • The page stated, without attribution, that the Sixth Circuit imposed $15,000 on each of two attorneys plus full payment of opposing counsel's fees, and that the brief had 24 fabricated citations. These figures come from a secondary source, Reason/Volokh, March 14, 2026. The primary sanctions order is not accessible to us: GovInfo has only the March 13, 2026 opinion in No. 25-5424, which does not mention AI or the amount, and CourtListener is closed to us. The page now attributes the figures to Reason/Volokh with the date, says costs rather than full fee reimbursement, and states that the order is not accessible. (1 page)

Source

Reason/Volokh, Lawyers citing nonexistent cases ordered to pay opponent's attorney fees, double costs, $15K fine (March 14, 2026)

Pages that carry this correction (1)

Whiting v. City of Athens and the Sixth Circuit page: no 'AI-fabricated' in the description, no 'most important circuit in America'

· 2 pages

What changed

  • The Whiting page's search-engine description called the citations AI-fabricated, but the Sixth Circuit asked whether generative AI was used, the attorneys did not answer and the court did not find that AI was used. The description now says the court ordered $15,000 from each of two attorneys plus double costs and fees over fake citations. The Sixth Circuit page opened with 'the most important circuit in America for AI sanctions law', a claim nothing supports; it now states what the opinion ordered. The internal source title that said '$30K+' now says $15,000 each plus double costs and fees. (2 pages)

Source

Whiting v. City of Athens, Nos. 24-5918/5919/25-5424 (6th Cir. Mar. 13, 2026), opinion PDF on Justia

Pages that carry this correction (2)

Whiting v. City of Athens: 13 pages that cite the case now match the opinion

· 13 pages

What changed

  • We checked every sentence on the pages that cite Whiting against the opinion itself. The court ordered $15,000 from each of two attorneys (paid to the court's registry), double costs and reimbursement of the appellees' attorneys' fees, for over two dozen fake citations and misrepresentations of fact; it did not give a combined figure above $30,000 or a count of 24. The court asked whether the briefs were written with generative AI, the attorneys did not answer, and the court did not find that AI was used. It sanctioned under Federal Rule of Appellate Procedure 38 and its inherent authority, not Rule 11, and the opinion states no disclosure duty, no fraud-on-the-court holding and no AI compliance framework. The Sixth Circuit pages and the Kentucky and Tennessee pages had said otherwise, and had called the case a precedent that district and state judges would follow or an enforcement bloc; those statements are removed or rewritten as our own advice. The opinion is marked recommended for publication. (13 pages)

Source

Whiting v. City of Athens, Nos. 24-5918/5919/25-5424 (6th Cir. Mar. 13, 2026), opinion PDF on Justia

Pages that carry this correction (13)

Westlaw, Lexis and vLex guide re-read against vendor pages: vLex 400+ courts, Key Number and KeyCite wording removed

· 4 pages

What changed

  • Removed the figure of 400+ U.S. courts with daily updates for vLex, the statement that Westlaw is known for the West Key Number System, and the wording that KeyCite checks whether an authority is still reliable. None of these appears on the vendor pages we re-read on October 28, 2026; Thomson Reuters describes KeyCite Overruling Risk as a way to make sure you cite reliable authority. The Westlaw Precision address now redirects to Westlaw Advantage, and the page names it. (1 page)
  • Removed the figure of 400+ US courts for vLex: vLex's coverage, Vincent AI and Fastcase Library pages, re-read on October 28, 2026, show more than one billion documents, 100+ countries and 14 languages, and no count of US courts. (2 pages)
  • Removed the statement that Westlaw users pay for both the database and the AI layer: no Thomson Reuters page we re-read on October 28, 2026 says how they are priced, so the page now says to ask for a written quote covering both. (1 page)

Source

Thomson Reuters, Westlaw (read October 28, 2026); Thomson Reuters, Westlaw Advantage (the Westlaw Precision address redirects here, read October 28, 2026); Thomson Reuters, CoCounsel Legal (read October 28, 2026); vLex, coverage (read October 28, 2026); vLex, Vincent AI (read October 28, 2026); vLex, Fastcase Library (read October 28, 2026)

Pages that carry this correction (4)

Seven 4th Circuit district pages and the AI disclosure index no longer say the Western District of North Carolina requires an AI certificate on every brief

· 8 pages

What changed

  • The Eastern, Middle and Southern pages for North Carolina and South Carolina, the Eastern and Western District of Virginia pages and both West Virginia pages said the Western District of North Carolina requires an AI certificate with every brief. That described its 2024 standing order. The court's July 29, 2026 order (No. 3:24-mc-104) superseded it and says a separate AI certificate is no longer required; every filing is instead deemed to certify evidentiary support, accurate authorities and verbatim quotations, with human verification. Body text, advice, FAQ answers and structured data now say that, and the claims that other 4th Circuit courts will follow were removed. (7 pages)
  • The AI disclosure index marked the Western District of North Carolina as REQUIRED and listed the District of Maryland with no label. Neither court requires AI disclosure: the Western District's order requires no separate certificate, and Maryland's Standing Order 2026-03 asks for verification of all filings without disclosure. Both now carry the PARTIAL label. (1 page)

Source

W.D.N.C., Order No. 3:24-mc-104, In re Use of Artificial Intelligence, July 29, 2026, paragraph 7 (scanned PDF, read October 28, 2026)

Pages that carry this correction (8)

Western District of North Carolina: the July 29, 2026 order replaced the 2024 AI certificate; District of Maryland: Standing Order 2026-03 added

· 2 pages

What changed

  • The Western District of North Carolina page described the 2024 standing order: a certificate filed with every brief saying either that no AI was used or that AI was used and verified. The court's July 29, 2026 order (No. 3:24-mc-104) superseded it and says a separate AI certificate is no longer required. Every filed document is now deemed to certify evidentiary support, warranted legal contentions, accurate authorities and verbatim quotations, with human verification by the signatory or a supervised attorney of record. The page's title block, body, advice, stat card and FAQ now say that, and add the order's evidence, discovery and sanctions provisions. (1 page)
  • The District of Maryland page said the court had no district-wide AI order. On August 5, 2026 the court issued Standing Order 2026-03: attorneys and unrepresented parties must verify the accuracy of all filings, especially when relying on AI, and fake or hallucinated citations or quotations can draw Rule 11 sanctions. It requires no AI disclosure. The page and the new court AI orders table now say so. (1 page)

Source

W.D.N.C., Order No. 3:24-mc-104, In re Use of Artificial Intelligence, July 29, 2026 (scanned PDF, 5 pages, read October 28, 2026); D. Md., Standing Order 2026-03, Misc. No. 00-308, August 5, 2026 (scanned PDF, 1 page, read October 28, 2026)

Pages that carry this correction (2)

vLex Vincent AI: 'over 40 state bars' replaced by vLex's own figure; Westlaw vs vLex citator verdict removed from the FAQ data

· 2 pages

What changed

  • The vLex Vincent AI page said Fastcase was free through 40+ state bars. We found no source for 40. vLex's own page says Fastcase Library is offered through 80+ state, county, and specialty bar associations, so the page now cites that figure with its scope and the date we read it. (1 page)
  • The structured data behind the Westlaw vs vLex vs Lexis FAQ said Westlaw's citator depth still outpaces vLex. We have not compared KeyCite with vLex's citation tools and found no source, so the answer now says we cannot say and tells readers to test both on the same citations. (1 page)

Source

vLex, Fastcase Library page (available through 80+ state, county, and specialty bar associations), read October 28, 2026; Thomson Reuters, Westlaw Advantage product page (no KeyCite comparison with vLex, no price), read October 28, 2026

Pages that carry this correction (2)

TAKE IT DOWN Act and xAI v. OpenAI: claims the primary sources do not support removed or corrected

· 2 pages

What changed

  • The TAKE IT DOWN Act page said the statute creates a rebuttable presumption about consent, cited the wrong Public Law number, gave a 20-year penalty and a $50,000 FTC penalty, and a 12-year sentence. The enacted text (Public Law 119-12) has no such presumption, sets up to 2 years for offenses involving adults and 3 years for minors, and gives no dollar figure for the FTC. The page also said the Act carves itself out of Section 230 and has a preemption clause; the text has neither, so those statements are replaced. (1 page)
  • The xAI v. OpenAI page gave a filing date, claims, damages, a page count and comparisons that we had not read in a primary source, and said the case was dismissed without noting leave to amend. It now follows OpenAI's motion to dismiss (three claims: federal trade secret misappropriation, intentional interference, unfair competition) and Reuters (dismissed with leave to amend, filed September 2025), and the unread figures and the Epic Games and Cisco comparisons are gone. (1 page)

Source

Public Law 119-12, TAKE IT DOWN Act (govinfo); OpenAI motion to dismiss, X.AI Corp. v. OpenAI, No. 3:25-cv-08133-RFL (N.D. Cal.), Dkt. 30 (October 2, 2025); Reuters via Yahoo Finance, Judge dismisses xAI trade secrets lawsuit against OpenAI (February 25, 2026)

Pages that carry this correction (2)

State-bar AI guidance counts (35, 14, 18, 30+, 42) replaced by a dated state-by-state count

· 5 pages

What changed

  • These pages gave a count of state bars with AI guidance (at least 35, at least 14, 18, 30+ and 42) with no source and no date, and the figures disagreed with each other. The count is now our own state-by-state index, read on October 28, 2026: 12 states and DC have a formal AI ethics opinion and 11 more states have other official guidance (some from state courts). The ABA's list of AI ethics opinions answered 403 that day and was not used. The claim that every state opinion puts competence and verification on the attorney is now limited to the four we re-read (Texas 705, Pennsylvania 2024-200, North Carolina 2024 FEO 1, D.C. 388), and Illinois is no longer named among the most detailed opinions because our index has it as not verified. (5 pages)

Source

AI Vortex state-by-state AI ethics index (13 jurisdictions with a formal opinion, 11 states with other official guidance), primaries re-read 2026-10-28; Texas Professional Ethics Committee Opinion 705; District of Columbia Bar Ethics Opinion 388

Pages that carry this correction (5)

"Non-delegable" wording removed from six pages

· 6 pages

What changed

  • Several pages called a lawyer's duty to verify, supervise or review "non-delegable", and one page said Judge Castel ruled in Mata v. Avianca that the duty to verify citations is non-delegable. The Whiting opinion does not use that word and we could not source it to Mata, so the pages now say the duty stays with the attorney and that Judge Castel found a Rule 11 violation for filing citations to cases that did not exist without checking them. (6 pages)

Source

Whiting v. City of Athens, Nos. 24-5918/5919/25-5424 (6th Cir. Mar. 13, 2026), opinion PDF on Justia

Pages that carry this correction (6)

Unsourced RELX and Microsoft market capitalizations and a Westlaw/Lexis per-user price range removed

· 9 pages

What changed

  • Removed market-capitalization figures we had no primary source for (RELX at $50B+ and Microsoft at $3 trillion+), together with the wording that leaned on them. The pages now name the vendor's parent company without a valuation. (9 pages)

Source

AI Vortex sourcing rule (a figure or claim on a page needs a published source; none found for the removed claim)

Pages that carry this correction (9)

Getty v. Stability AI (UK): judgment date, who is appealing and what the court decided corrected against the judgment

· 4 pages

What changed

  • Corrected the UK judgment date to 4 November 2025, and the appeal: Getty was granted permission to appeal on 16 December 2025, and Stability was refused permission on the trade mark findings. Removed a parameter-count figure and a quotation that the judgment does not contain. (1 page)
  • Rewrote what the court decided: Getty dropped its training, outputs and database right claims before closing, and the court dismissed only the secondary infringement claim (a model that stores no copies is not an infringing copy under UK law) and found trade mark infringement for a few watermark examples. The court made no finding that training or scraping was lawful or unlawful. (1 page)
  • Corrected what this page said about the UK ruling: it was handed down in November 2025 (not February 2025), and the court made no finding that scraping or training copying was infringing, because Getty dropped its training claim before closing. (3 pages)
  • The Andersen v. Stability AI FAQ said the Ninth Circuit's rulings in cases such as Getty Images were likely to influence the outcome. Getty Images v. Stability AI is a UK case, decided by the English High Court on 4 November 2025, and it does not bind US courts; the FAQ now says so. Its status line was also refreshed from April 2026 to October 24, 2026. (1 page)

Source

Getty Images (US) Inc & Ors v Stability AI Ltd [2025] EWHC 2863 (Ch), 4 November 2025; Getty Images (US) Inc & Ors v Stability AI Ltd (Re Form of Order) [2025] EWHC 3343 (Ch), 16 December 2025

Pages that carry this correction (4)

Sequoia $60 billion legal figure and Harvey $1B+ raised removed from the remaining pages

· 7 pages

What changed

  • Removed the $60 billion legal figure that these pages attributed to Sequoia. The essay text lists legal transactional work at $20-25 billion and carries no larger legal total, so the pages now use that figure. (6 pages)
  • Removed Harvey's "$1B+ raised": Harvey's own announcements state no total raised, so the page now cites only the $15.5 billion valuation from its September 9, 2026 announcement. (1 page)

Source

Sequoia Capital, Services: The New Software (March 5, 2026); Harvey, $550M round announcement (September 9, 2026); Harvey, $200M round announcement (March 25, 2026)

Pages that carry this correction (7)

Sequoia legal figures, Harvey revenue and funding claims, Spellbook funding and Clio valuation corrected or attributed against primaries

· 8 pages

What changed

  • Removed the $36 billion paralegal line, the $60 billion legal total and the $15-20 billion patent line attributed to Sequoia: Sequoia's essay text lists legal transactional work at $20-25 billion and none of the others. The priority-map graphic is not part of the essay text and is not covered here. (2 pages)
  • Removed Harvey's "$100M+ in ARR" and the per-seat price derived from it: Harvey's own announcements state no revenue figure. (1 page)
  • Replaced Spellbook's outdated "$20M+ in funding" with its $50 million Series B at a $350 million post-money valuation (October 2025), attributed to BetaKit. (1 page)
  • Removed "$1B+ raised from Sequoia and GIC" for Harvey: its September 9, 2026 announcement names Sequoia and GIC as existing investors and gives no such total. (1 page)
  • Removed the unsupported "best-funded legal AI company" superlative for Harvey; its round sizes are unchanged. (1 page)
  • Attributed Clio's valuation and Series G figures to LawNext (clio.com returned 403 and was not read). (2 pages)

Source

Sequoia Capital, Services: The New Software (March 5, 2026); Harvey, $550M round announcement (September 9, 2026); Harvey, $200M round announcement (March 25, 2026); BetaKit, Spellbook Series B (October 9, 2025); LawNext, Clio completes vLex acquisition and announces Series G (November 10, 2025)

Pages that carry this correction (8)

Expired status sweep: Italian DPA appeal, Musk v. OpenAI trial, CoCounsel general availability, NIST RMF 1.1, NY Part 161 now read as of October 28, 2026

· 7 pages

What changed

  • The Italian DPA v. OpenAI page said OpenAI's appeal was pending and the fine stood. The Garante's own page says the Tribunal of Rome accepted the appeal in judgment No. 4153/2026 (March 18, 2026) and removed the decision from its site. The page now says so, with the date. (1 page)
  • The Musk v. OpenAI page was written ten days before trial and still read as if it were upcoming. It now carries a status dated October 28, 2026: per Reuters, the jury found on May 18, 2026 that Musk filed too late, and Musk said he would appeal. The docket gave us no access (403), so the status rests on press reports. (1 page)
  • The CoCounsel and Claude page said the next generation of CoCounsel Legal was expected to reach general availability in summer 2026. Thomson Reuters announced it available in the United States on August 20, 2026, with Canada, the UK and Australia expected later this year. (1 page)
  • The compliance monitoring guide said NIST released an AI RMF Profile for critical infrastructure in April 2026 and would publish RMF 1.1 addenda. NIST published a concept note on April 7, 2026, says AI RMF 1.0 is being revised, and lists no 1.1. The page now says that. (1 page)
  • Removed two unsourced forecasts that a district would adopt an AI policy (Southern District of Texas, Eastern District of New York). We could not confirm either, so the pages now send readers to the current local rules. (2 pages)
  • The New York page said Part 161 takes effect in June 2026. It took effect on June 1, 2026, per the New York State Bar Association; the page now says so. (1 page)

Source

Garante per la protezione dei dati personali, page on the OpenAI decision (notes Tribunal of Rome judgment No. 4153/2026, published March 18, 2026, accepted OpenAI's appeal), read October 28, 2026; Thomson Reuters press release, August 20, 2026: next generation of CoCounsel Legal available now in the United States, read October 28, 2026; NIST, AI Risk Management Framework page (April 7, 2026 concept note; AI RMF 1.0 being revised; no 1.1 listed), read October 28, 2026; Al Jazeera (Reuters report), Musk loses lawsuit against OpenAI, May 18, 2026 (secondary: the CourtListener docket returned 403), read October 28, 2026; New York State Bar Association, June 2, 2026: Part 161 effective June 1, 2026 (secondary: nycourts.gov returned 403), read October 28, 2026

Pages that carry this correction (7)

Court quotations we could not find in a primary source removed from three case pages

· 3 pages

What changed

  • The AI sanctions tracker page described an Eastern District of Virginia sentencing memorandum, a named judge's quoted order and a DOJ reprimand. We could not trace any of it to a court record, so it is gone. The section now covers a July 2026 Western District of Michigan order, as reported by Bloomberg Law. (1 page)
  • The TAKE IT DOWN Act page attributed two quotations and evidentiary rulings to a named judge in the first conviction under the Act. We could not find them in any court record, so they are removed; the section now reports only what Spectrum News and the U.S. Attorney's Office said. (1 page)
  • The xAI v. OpenAI page named the judge as James Lin and quoted three passages from her order. The judge is Rita F. Lin, and the quotations do not appear in the coverage we read, so they are replaced by a paraphrase attributed to Reuters. (1 page)

Source

Bloomberg Law, DOJ Called Out by Judge for Alleged AI Use, Fake Case Citation (July 16, 2026); Spectrum News, Take It Down Act conviction (April 8, 2026); Reuters via Yahoo Finance, Judge dismisses xAI trade secrets lawsuit against OpenAI (February 25, 2026)

Pages that carry this correction (3)

Colorado AI Act: pages that said it takes effect in June 2026 now say SB 26-189 replaced it and applies from January 1, 2027

· 12 pages

What changed

  • Removed the statement that the Colorado AI Act takes effect in June 2026 (or on June 1, 2026, or in February 2026) and the wrong bill number SB 21-169. Colorado's General Assembly shows SB 24-205 was repealed and reenacted by SB 26-189, signed May 14, 2026, a narrower automated decision-making law that applies to consequential decisions made on or after January 1, 2027. The pages now say so, and describe its notice, human-review and developer-documentation duties instead of the 2024 impact-assessment duties. (12 pages)
  • Removed the claim that the Colorado Attorney General signalled a phased enforcement approach from June 2026 to June 2027: we found no source. SB 26-189 requires 60 days' notice and a chance to cure before any action before January 1, 2030. (1 page)

Source

Colorado General Assembly, SB26-189, signed act of May 14, 2026 (section 5: takes effect and applies to consequential decisions made on or after January 1, 2027), read October 28, 2026; Colorado General Assembly, SB24-205 (the 2024 act, repealed and reenacted by SB26-189), read October 28, 2026

Pages that carry this correction (12)

A&O Shearman agentic playbook: unsourced $50,000-150,000 rollout budget removed

· 2 pages

What changed

  • Removed a $50,000-150,000 budget range for the rollout phase that we could not source. The wording now tells readers to build the rollout budget from their own training plan and staff time. (2 pages)

Source

AI Vortex sourcing rule (a figure or claim on a page needs a published source; none found for the removed claim)

Pages that carry this correction (2)

Westlaw vs Lexis vs vLex guide: unsourced coverage, AI, pricing and free-trial claims reformulated

· 1 page

What changed

  • Removed the claim that vLex is the one platform with a free trial. Thomson Reuters' Westlaw Precision and CoCounsel Legal pages show a free-trial request and LexisNexis offers a 2-day free trial of Lexis+ with Protégé. The page now lists all three, with the date read. (1 page)
  • Removed rankings of coverage and the founding date given for the Key Number System: we found no primary source for them and have not benchmarked the platforms. Coverage is now described as each vendor describes it, with the date we read it. (1 page)
  • Removed unsourced statements about contract terms, what firms pay, how many attorneys use the AI, vLex secondary sources and Lexis+ AI limits. The page now says we have no such data and points to written quotes. (1 page)

Source

Thomson Reuters, Westlaw (read October 26, 2026); Thomson Reuters, Westlaw Precision, free trial (read October 26, 2026); LexisNexis, Lexis+ with Protégé, 2-day free trial and pricing (read October 26, 2026); vLex, coverage (read October 26, 2026)

Pages that carry this correction (1)

Westlaw, Lexis and vLex pages: unsourced rankings, citator, Key Number, market and free-trial claims reformulated

· 9 pages

What changed

  • Removed rankings of Westlaw, Lexis and vLex on case law, secondary sources, citators (KeyCite and Shepard's), the Key Number System and Practical Law, including a claim of independent studies and a count of key numbers. We found no primary source for them and have not benchmarked the platforms. Each page now says what the vendor says on its own page, with the date we read it, and says where we have not compared. (7 pages)
  • Removed the claim that the Westlaw-Lexis duopoly controlled legal research for 50 years, revenue and per-user price figures for Thomson Reuters and RELX, claims about vendor discounts and contract terms, and the statement that Lexis has overtaken Westlaw on AI. None had a source on the page. The pages now say neither vendor publishes a firm price and point to written quotes. (2 pages)
  • Removed unsourced claims about vLex: that it is the most undervalued platform, covers 85% of research needs, has no per-search fees, lacks US secondary sources, trails on citators, and holds international content Westlaw and Lexis lack. vLex now appears with its own figures (more than one billion documents, 100+ countries, 14 languages) and the bar-association page, with the date read. (1 page)
  • Removed the claims that CoCounsel can only be trialled through a Westlaw subscription, that the trial is a sales-led pilot, that promotional trials of 30-60 days exist unadvertised, and that CoCounsel is the most capable legal AI assistant. Thomson Reuters' CoCounsel Legal page shows a "Request free trial" button; the trial's length, limits and eligibility are not stated on the page, and the pages now say so. (3 pages)

Source

Thomson Reuters, Westlaw (read October 26, 2026); Thomson Reuters, CoCounsel Legal, free trial (read October 26, 2026); LexisNexis, Lexis+ with Protégé (read October 26, 2026); vLex, Vincent AI (read October 26, 2026); vLex, coverage (read October 26, 2026); vLex, bar associations (read October 26, 2026)

Pages that carry this correction (9)

Westlaw AI-Assisted Research no longer labelled 'Most Accurate Overall' on the legal research ranking

· 1 page

What changed

  • Removed the heading and top-pick wording that called Westlaw AI-Assisted Research the most accurate tool. The only independent test we cite, the Stanford RegLab and HAI study of 2024 versions, found Lexis+ AI accurate on 65% of queries and Westlaw AI-Assisted Research on 42%. The page now ranks Lexis+ with Protégé first on that evidence, says the figures are old, and notes that CoCounsel and Vincent were not tested. (1 page)

Source

Magesh et al., Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools (Stanford RegLab and HAI, first released May 2024)

Pages that carry this correction (1)

Kira: field count unified at 1,400+ pre-built smart fields, attributed to Litera

· 6 pages

What changed

  • Six pages said Kira handles 1,000+ legal concepts, data points, provision types or extraction models, while another of our pages said 1,400+. Litera's own Kira page says 1,400+ pre-built smart fields, plus fields you train yourself, and 95%+ precision on target provisions. All pages now give the 1,400+ figure as Litera's, with the date we read it. We have not tested Kira's accuracy. (6 pages)

Source

Litera, Kira (read October 27, 2026)

Pages that carry this correction (6)

CoCounsel described as quote-only or sold by quote as a whole product, corrected to Thomson Reuters' in-house starting price

· 8 pages

What changed

  • These pages said CoCounsel as a whole is quote-only or sold by quote. Thomson Reuters lists CoCounsel from $340.85 per month for in-house teams (Essentials; $882.30 and $1,694.90 for the other listed plans), and the page loads. Law-firm plans and the Westlaw Precision + CoCounsel bundle are quoted, so the pages now say that and give the read date. Harvey, Spellbook, Lexis+ AI and Everlaw still show no price on their own sites when we read them, so their quote-only wording is unchanged. (8 pages)

Source

Thomson Reuters, CoCounsel plans for in-house legal teams (read October 27, 2026)

Pages that carry this correction (8)

Claims that Spellbook Library or legal AI vendors are trained or fine-tuned on a firm's or legal data, without a primary source, retired or attributed, ninth pass

· 12 pages

What changed

  • Rewrote the Spellbook Library description. Removed statements that Library trains a model on a firm's contracts, that it is the first feature of its kind, that it improves after months of use, that setup takes weeks, that it pays back within months, that it shipped with the Series B (it launched July 31, 2025) and that firms signing two-year terms are in effect signing longer ones. Spellbook's own launch post says Library indexes the documents a firm connects or uploads, lets a lawyer search past clauses from Word and revises the inserted clause to fit the new agreement; it gives no setup time, payback or measure of improvement, and the pages now say that, with the date we read it. (7 pages)
  • Removed statements that CoCounsel, Vincent, Harvey, Spellbook, Kira, Luminance or Relativity models are fine-tuned or trained on legal data or on a firm's work. Thomson Reuters describes CoCounsel as AI grounded in Westlaw and Practical Law content, vLex says Vincent is built on its legal database, Harvey says it builds a bespoke model only if a firm asks, and Relativity says aiR work product is grounded in the system of record; none of them publishes how its models are trained, and the pages now say that, with the date we read it. (5 pages)

Source

Spellbook, Introducing Library and Smart Clause Drafting (read October 25, 2026); Spellbook, home page (read October 25, 2026); Thomson Reuters, CoCounsel Legal (read October 25, 2026); vLex, Vincent AI (read October 25, 2026); Harvey, security page (read October 25, 2026); Relativity, aiR (read October 25, 2026)

Pages that carry this correction (12)

Relativity, Everlaw, Luminance and Legora security, adoption and price claims without a primary retired or attributed

· 21 pages

What changed

  • Replaced our Relativity user and Am Law figures with what Relativity's own security page says: 300K+ legal professionals, 30+ countries and 193 of the Am Law 200 relying on it. We had printed 300,000 users in 40+ countries and 198 of the Am Law 200, which do not match that page, and we dropped the ranking language that had no source. (4 pages)
  • Relativity lists ISO/IEC 27001:2022 and 27018, SOC 2 Type II, SOC 3 and HIPAA compliance for RelativityOne, and says its FedRAMP Moderate authorization applies to Relativity aiR Government. The pages no longer call that the highest certification profile in legal AI or say FedRAMP is unique to Relativity (Everlaw also lists it), and the unsourced $3.6B valuation and the $100M+ raised figure for Luminance are gone. (5 pages)
  • Removed the claims that Everlaw has stronger government adoption than Relativity, that named agencies use it and that it is the fastest-growing e-discovery challenger. Everlaw's security page lists SOC 2 Type 2, FedRAMP Moderate, GovRAMP Moderate and ISO/IEC 27001:2022 and names no agency customers, so the pages now say only that. (4 pages)
  • Removed the statement that Luminance offers on-premise deployment, which its security page does not list. That page lists ISO 27001 and SOC 2 Type II and describes AWS hosting with AES-256 encryption, so the pages tell readers to ask Luminance about hosting and data residency. Also removed the claim that a Luminance team pays an order of magnitude more than Claude (Luminance lists no price), the unsourced claim that Kira and Luminance offer air-gapped deployment, and the label that named Kira as a SymphonyAI product (Litera owns Kira). (11 pages)
  • Removed an inferred Legora price range and an inferred pricing strategy: legora.com has no pricing page and the newsroom lists no price, so the page says no price was found. Legora's own announcement of October 21, 2026 gives more than 80 markets and offices in North America, Europe and Asia-Pacific, which replaces the 16-country figure. The Lexis+ AI per-seat range attributed to unnamed third parties is also gone: LexisNexis publishes no list price and we found no named source. (2 pages)

Source

Relativity, RelativityOne security (read October 26, 2026); Everlaw, security (read October 26, 2026); Luminance, security (read October 26, 2026); Legora, newsroom: Series E announcement (read October 26, 2026)

Pages that carry this correction (21)

Spellbook, ChatGPT Team and practice-management prices and Harvey and CoCounsel 'pre-loaded' claims without a readable primary retired or attributed

· 13 pages

What changed

  • Removed the per-seat monthly price we had given for Spellbook. Spellbook's pricing page says pricing is determined by the number of team members on a license and points to a demo for customized pricing; it publishes no plan names or per-seat amounts. The pages now say it is priced by quote. (6 pages)
  • Re-read the PracticePanther and Rocket Matter pricing pages. PracticePanther lists $59, $79 and $99 per user per month on monthly billing and $49, $69 and $89 billed annually; the page no longer says its Business plan includes AI or that support staff cost less, because the pricing page says neither. Rocket Matter's Premier total was recalculated for 8 users and the Ultimate tier, which its page does not list, was removed. (1 page)
  • Removed the statements that Harvey comes with its own built-in legal corpora and that CoCounsel's litigation templates are built in. Harvey says it pulls from documents, data, Ask LexisNexis and 500+ regional knowledge sources; Thomson Reuters says CoCounsel Legal is grounded in Westlaw and Practical Law content. Neither says which templates or corpora ship built in. (2 pages)
  • Removed the $25 per user per month price for ChatGPT Team on the employment contract review and four document drafting pages. OpenAI's pricing and help pages could not be read on October 26, 2026 (the sites refused the request), so the pages point to OpenAI's pricing page instead, and the combined $50 a month comparison built on that price is gone. (5 pages)

Source

Spellbook, pricing (read October 26, 2026); Harvey, platform (read October 26, 2026); Thomson Reuters, CoCounsel Legal (read October 26, 2026); PracticePanther, pricing (read October 26, 2026); Rocket Matter, pricing (read October 26, 2026)

Pages that carry this correction (13)

Spellbook usage thresholds, payback, price estimates, deployment times and training claims without a primary source, retired or attributed, tenth pass

· 10 pages

What changed

  • Removed the volume thresholds, the history length and the payback period we had put on Spellbook. Spellbook's pricing page says pricing is custom, based on the number of team members on a license, and that teams can try it free for 7 days; it publishes no price, no minimum volume, no payback period and no history requirement, and the pages now say that, with the date we read it. (6 pages)
  • Removed seat-price estimates and cost-per-review figures for Spellbook that came from unnamed secondary coverage. Spellbook publishes no per-seat price. (2 pages)
  • Removed deployment times, procurement times and the claim that the free trial needs no payment card. Spellbook's pricing page describes the Suite as a Word add-in and the 7-day trial; it gives no deployment or procurement time and does not say whether a card is required. (6 pages)
  • Removed statements that Library trains on a firm's contracts, improves with history, compares incoming contracts against it or surfaces negotiation positions, and that Spellbook's models are English-trained or playbook-trained. Spellbook's Library post says Library indexes the documents a firm connects or uploads so a lawyer can search past clauses from Word; it does not say a model is trained on them, and Spellbook does not publish its language coverage. (9 pages)

Source

Spellbook, pricing page (read October 26, 2026); Spellbook, Introducing Library and Smart Clause Drafting (read October 26, 2026); Spellbook, home page (read October 26, 2026)

Pages that carry this correction (10)

Spellbook and the Canadian Bar Association: student-training and membership claims without a source retired

· 11 pages

What changed

  • Removed the statements that Canadian law students would learn the tool through CBA membership, that hiring firms would inherit staff already fluent in it and that this creates switching costs for years. None of the sources we read describes a student program; Spellbook's CBA page says members get discounts on annual plans. What is left is marked as our inference or as an open question. (5 pages)
  • Replaced the list of member categories the deal covers with BetaKit's wording, "legal professionals". The BusinessWire release is not quoted here, so the membership categories and any claim about Quebec notaries being in scope are no longer stated. (10 pages)

Source

Spellbook, CBA members page (read October 26, 2026); BetaKit, Spellbook partners with Canadian Bar Association (read October 26, 2026); CP24, Canadian Bar Association signs two-year deal with Spellbook (read October 26, 2026)

Pages that carry this correction (11)

Spellbook capabilities, prices and contract terms without a primary source retired or attributed, eleventh pass

· 4 pages

What changed

  • Removed the statement that Spellbook tracks a counterparty's past markups, recognizes their patterns and surfaces likely concession points. Spellbook's own pages describe redlining with every suggestion tracked; none describes a counterparty-history feature. (1 page)
  • Removed the claim that Spellbook ships with ready-made contract document sets and checklists tied to each contract type. Spellbook says it compares contracts to market standards using realtime market data and lets a team build playbooks with rules and fallbacks; it does not publish which documents or contract types sit behind either. (2 pages)
  • Removed statements that Spellbook learns a firm's preferred language, remembers its standards in a playbook that never resets, compares incoming contracts against it automatically and wins on passive review. The pages now use Spellbook's own wording: Library indexes the documents a firm connects, and playbooks turn a team's standards into automated checklists. (4 pages)
  • Removed the $99 to $165 per user per month range, attributed to unnamed user reports, and the statement that Spellbook has no seat minimums or multi-year commitments. Spellbook's pricing page says pricing is determined by the number of team members on a license, mentions a 7-day trial and lists no dollar figures or commitment terms. (1 page)

Source

Spellbook, home page (read October 26, 2026); Spellbook, pricing page (read October 26, 2026); Spellbook, Benchmarks (read October 26, 2026); Spellbook, Playbooks (read October 26, 2026)

Pages that carry this correction (4)

Claude Team and ChatGPT Business per-seat prices without a readable primary attributed or removed

· 18 pages

What changed

  • Gave the Claude Team price with its billing term and source. Anthropic's pricing page lists the Team standard seat at $25 per seat per month billed monthly and $20 billed annually (premium seats $125 monthly and $100 annual); pages that showed a bare $25 now say which term it is and the date we read it. A Claude Team Standard total that applied $25 to annual billing was corrected, an Enterprise price of $30 per seat that the page does not show was replaced with what Anthropic lists (US$20 per seat plus usage at API rates, 20-seat minimum), and a line that called a $25 plan a free tier was reworded. (14 pages)
  • Removed the per-user monthly price we had given for ChatGPT Team and Business, and the monthly price we had given for ChatGPT Plus. OpenAI's pricing page and help center could not be read when we checked them on October 26, 2026, so those figures could not be confirmed with OpenAI; the pages now point to OpenAI's pricing page instead of quoting one. (7 pages)

Source

Anthropic, Claude pricing (read October 26, 2026)

Pages that carry this correction (18)

Unsourced count of federal judges with AI orders removed from 88 more pages

· 88 pages

What changed

  • This page gave a national count of federal judges with AI orders, a figure of more than 300 that we could not trace to a source. No source gives an authoritative national count; the page no longer states one. Individual judges set their own requirements, so check the assigned judge's orders before filing. The same correction was made on How Many Federal Judges Require AI Disclosure. (88 pages)

Source

What Courts Require on AI Disclosure (our page: no authoritative national count; Ropes & Gray tracker scope); How Many Federal Judges Require AI Disclosure (our page, corrected October 26, 2026)

Pages that carry this correction (88)

Unsourced count of federal judges with AI orders removed from 40 pages

· 40 pages

What changed

  • This page gave a national count of federal judges with AI orders, a figure of more than 300 that we could not trace to a source (some passages attributed it to the Ropes & Gray or Bloomberg Law trackers, but neither is a count of judges, and we have not audited them). No source gives an authoritative national count; the page no longer states one. Individual judges set their own requirements, so check the assigned judge's orders before filing. The same correction was made on How Many Federal Judges Require AI Disclosure. (40 pages)

Source

What Courts Require on AI Disclosure (our page: no authoritative national count; Ropes & Gray tracker scope); How Many Federal Judges Require AI Disclosure (our page, corrected October 26, 2026)

Pages that carry this correction (40)

How many federal judges require AI disclosure: no authoritative count, Mata sanction restated

· 1 page

What changed

  • This page said that more than 300 federal judges require AI disclosure, that 13 district courts had court-wide rules, and that the number of judges grew from 50 to 150 to 300 after Mata. We could not trace those figures to a source, and the page's own search description said 60. No source gives an authoritative national count; the page now says so and points to our judge profiles (91 judges, checked October 21, 2026) and to the tracker explanation on What Courts Require on AI Disclosure. It also said the Mata attorneys faced $5,000 fines each and mandatory legal education; the court imposed $5,000 jointly and severally on two lawyers and their firm and ordered notification letters (our Mata case brief, read October 26, 2026). Other pages still repeat a 300+ judge figure and are queued for the same correction. (1 page)

Source

What Courts Require on AI Disclosure (our page: no authoritative national count; Ropes & Gray tracker scope); Mata v. Avianca, Opinion and Order, June 22, 2023, as summarized in our case brief

Pages that carry this correction (1)

CoCounsel described as having no published price at the tier level, corrected to Thomson Reuters' in-house starting price

· 10 pages

What changed

  • These pages said Thomson Reuters publishes no price for CoCounsel tiers, that the tiers are quote-only, or that its pricing page could not be read. Thomson Reuters lists CoCounsel from $340.85 per month for in-house teams (Essentials; $882.30 and $1,694.90 for the other listed plans), and the page loads. Firm plans and the Westlaw Precision + CoCounsel bundle are quoted, so the pages now say that and give the read date. Harvey, Spellbook, Lexis+ AI and Westlaw Advantage still show no price on their own sites when we read them, so their quote-only wording is unchanged. (10 pages)

Source

Thomson Reuters, CoCounsel plans for in-house legal teams (read October 26, 2026)

Pages that carry this correction (10)

Vendor claims about learning from a firm's own work, without a primary source, retired or attributed, eighth pass (Harvey and Spellbook)

· 14 pages

What changed

  • Removed statements that Harvey trains on a firm's own data as a standard feature, that it learns a firm's templates or claim construction arguments, and that its model carries legal-domain tuning that produces better drafts. Harvey says it never trains on customer data by default and builds a bespoke model for a firm only if the firm explicitly asks (security page), and its platform page says Workflows can be tailored to a firm's needs; the pages now say that, with the date we read it, and we did not test whether a bespoke model learns a firm's templates. Harvey does not publish how its models are trained and we found no independent benchmark. (11 pages)
  • Removed statements that Spellbook learns from a firm's precedent contracts, improves with months of use or trains on the firm's prior contracts. Spellbook's own Library announcement says the Library indexes the documents a firm connects, lets a lawyer search past clauses and adapts the inserted clause to the new agreement; the pages now say that, with the date we read it, and we found no published measure of improvement with use. (3 pages)

Source

Harvey, security page (read October 25, 2026); Harvey, platform page (read October 25, 2026); Spellbook, Introducing Library and Smart Clause Drafting (read October 25, 2026)

Pages that carry this correction (14)

Harvey claims about training on a firm's own data, without a primary source, retired, seventh pass (Harvey comparisons, use cases, agentic pages, pricing)

· 12 pages

What changed

  • Removed statements that Harvey comes with training on a firm's own data as a standard feature, that its models learn a firm's templates and work product over time, and that its workflows are custom-trained. Harvey says it never trains on customer data by default and builds a bespoke model for a firm only if the firm explicitly asks; the pages now say that, with the date we read it, and we did not test whether a bespoke model learns a firm's templates. Where a page named Harvey's workflow features, it now cites Harvey's platform page, which says Workflows can be tailored to a firm's needs. (12 pages)

Source

Harvey, security page (read October 25, 2026); Harvey, platform page (read October 25, 2026)

Pages that carry this correction (12)

Harvey firm-specific training, Spellbook training and generic purpose-built wording without a primary source retired, sixth pass (Harvey vs ChatGPT, Harvey guides, contract drafting workflow)

· 7 pages

What changed

  • Removed statements that Harvey comes with training specific to each firm, that large firms typically run custom Harvey instances, and that this training locks buyers in. Harvey says it never trains on customer data by default and builds a bespoke model exclusively for a firm only if the firm asks; the pages now say that, with the date we read it, and the lock-in sentence now says switching cost depends on the contract. (6 pages)
  • Replaced a statement about what Spellbook's models learn from and a generic claim about dedicated tools at each drafting step. Spellbook describes itself as an AI contract review platform for transactional lawyers and does not publish its training data. (1 page)

Source

Harvey, security page (read October 25, 2026); Harvey, platform page (read October 25, 2026); Spellbook, home page (read October 25, 2026)

Pages that carry this correction (7)

Vendor training, purpose-built and price wording without a primary source retired, fifth pass (DISCO Cecilia, Spellbook, Harvey, Luminance) and an unsourced damages figure removed from Concord page tags

· 5 pages

What changed

  • Removed statements that Harvey runs on fine-tuned models trained on legal documents, that Spellbook is purpose built, and that DISCO's Cecilia was designed from scratch for million-document datasets. Harvey says it never trains on customer data by default and builds a bespoke model only on request; Spellbook does not publish how its models are trained; DISCO describes Cecilia as agentic AI purpose built for complex cases and queries across massive datasets, with no document-count claim. The pages now say that, with the date we read it. (3 pages)
  • Replaced the page's search and share description, which put an enterprise price on Luminance and ranked Claude's reasoning above it on novel contracts. Luminance lists no price and neither vendor published a benchmark against the other; Claude Pro is $20 per month billed monthly. The description now says only that. (1 page)
  • Removed an over-precise damages figure from the page's article tags. Press reports say Concord II seeks more than $3 billion in statutory damages; no primary source gives a more exact number, and the body text already says more than $3 billion. (1 page)

Source

DISCO, Cecilia page (read October 25, 2026); Harvey, security page (read October 25, 2026); Spellbook, home page (read October 25, 2026); Luminance, Luna Crescent page (read October 25, 2026); Anthropic, Claude pricing (read October 25, 2026)

Pages that carry this correction (5)

Vendor training and purpose-built claims without a primary source retired, fourth pass (compare and guide pages)

· 8 pages

What changed

  • Removed statements that Spellbook and Harvey models are trained on legal contracts or legal data, that Harvey offers firm-specific training as a matter of course, and that Spellbook, Luminance, CoCounsel and Harvey are purpose built. Spellbook does not publish how its models are trained; Harvey says it never trains on customer data by default and builds a bespoke model only on request; Luminance describes its Luna Crescent model for specialist legal tasks; Thomson Reuters describes CoCounsel as AI grounded in Westlaw and Practical Law content. The pages now say that, with the date we read it. (8 pages)

Source

Harvey, platform page (read October 25, 2026); Harvey, security page (read October 25, 2026); Spellbook, home page (read October 25, 2026); Luminance, Luna Crescent page (read October 25, 2026); Thomson Reuters, CoCounsel page (read October 25, 2026)

Pages that carry this correction (8)

Vendor training, purpose-built, accuracy and language-count claims without a primary source retired, third pass

· 37 pages

What changed

  • Removed statements that Harvey's models are trained or fine-tuned on legal material, or trained per firm. Harvey describes its platform as purpose built tools for legal work and does not publish how its models are trained; its security page says it never trains on customer data by default and builds a bespoke model only for a firm that asks. The pages now say that, with the date we read it. Spellbook describes itself as an AI contract review platform for transactional lawyers and does not publish its training data. (19 pages)
  • Removed statements that Luminance models were built from scratch or trained only on legal documents. Luminance says its proprietary model, Luna Crescent, starts from a general-purpose base model, is trained on Luminance contract data and runs next to third-party models. (4 pages)
  • Removed Luminance language counts (60, 14 and 80 languages). Its product pages publish no count; the only figure we found is a 2022 Luminance blog post that says over 80 languages. The pages now say no count is published and to ask for the supported-language list. (15 pages)
  • Removed figures with no primary: an 80% share of analytical work for Claude Pro, Luminance review times, 90%+ accuracy figures for Blue J, Kira, Luminance and Harvey, and a 300,000-decision count for Blue J. The pages say we have no measured figure or cite the current primary. (9 pages)

Source

Harvey, platform page (read October 25, 2026); Harvey, security page (read October 25, 2026); Luminance, Luna Crescent page (read October 25, 2026); Luminance, blog post of October 28, 2022 on central and eastern Europe; Spellbook, home page (read October 25, 2026); Blue J, home page (read October 25, 2026)

Pages that carry this correction (37)

Vendor 'trained on legal data' and 'purpose-built' claims without a primary source retired, second pass

· 10 pages

What changed

  • Removed statements that Harvey, Luminance or Spellbook run on models trained only on legal data. None of them publishes training details; Spellbook's own page says it uses GPT-5, Claude and other LLMs. Also removed a '60+ languages' count for Luminance that its page does not state. The pages now quote what each vendor says, with the date we read it. (6 pages)
  • Removed 'purpose-built' as our own verdict on Legora, Briefpoint, Spellbook and CoCounsel, and an American-law training claim about Harvey. The pages now attribute self-description to the vendor with URL and date, or say we did not test it. (4 pages)
  • Removed an 80/20 split of lawyers' work between general-purpose AI and Harvey, and the 'factor of 50' overpayment claim: neither has a source. The page now says no measured share exists and to test both tools on your own matters. (1 page)

Source

Harvey, platform page (read October 25, 2026); Spellbook, product page (read October 25, 2026); Luminance, contract intelligence page (read October 25, 2026); Briefpoint, product page (read October 25, 2026)

Pages that carry this correction (10)

Vendor 'purpose-built' and training claims without a primary source retired

· 9 pages

What changed

  • Removed statements that Harvey was trained or fine-tuned on legal data, and a ranking of it as the most sophisticated platform. Harvey's own page calls its agents purpose built and publishes no training details, so the pages now attribute that wording to Harvey with the date we read it. (5 pages)
  • Removed claims that Luminance's models perform better on contracts, that Spellbook's suggestions come from training on millions of agreements, and that Relativity aiR was made for one kind of work, none of which the vendors publish. The pages now quote what each vendor says on its own page. (4 pages)
  • Removed a market-share figure, a claim that no vendor serves small firms, and head-to-head quality claims about Claude and Harvey that we had not tested. The pages now say what we did not test. (3 pages)

Source

Harvey, platform page (read October 25, 2026); Spellbook, product page (read October 25, 2026); Luminance, contract intelligence page (read October 25, 2026); Relativity, aiR page (read October 25, 2026); Briefpoint, product page (read October 25, 2026)

Pages that carry this correction (9)

Claims about Kira and Clio Duo corrected or attributed: Litera's own pages read, Clio's pages unreadable, second pass

· 15 pages

What changed

  • Removed Kira adoption percentages and leader or accuracy-leader superlatives. Litera's Kira page gives no Kira adoption figure and there is no independent accuracy ranking. (7 pages)
  • Removed a July 2025 launch date and free-of-charge claim for generative AI in Kira. Litera says Lito is included for Kira customers at no additional charge, and that Kira Grid Chat is available now. (1 page)
  • Removed Kira price ranges and pricing bases. Litera publishes no Kira price and points to Request a Demo. (5 pages)
  • Removed named-firm adoption, a deal anecdote, time-saved and deployment-time figures, document-system integrations and firm-size positioning for Kira that had no source. Litera lists integrations with Microsoft 365, Litera Transact, HighQ and Intralinks. (6 pages)
  • Replaced statements that Kira and Clio Duo are SOC 2 compliant with a note to ask each vendor for its current security documents. Clio's site was not available for checking, so no Clio security page was checked. (8 pages)
  • Clio AI pricing had been stated in incompatible ways. Pages now say Clio states AI is included on Core and above and tell readers to confirm with Clio; derived prices, time-saved and ROI figures and a launch year were removed. Clio's pricing pages were not available for checking. (10 pages)

Source

Litera, Kira page (read October 25, 2026); Litera, our story (read October 25, 2026); Litera, Kira Grid Chat blog post of July 17, 2026 (read October 25, 2026); Law Society Gazette, August 19, 2021, on the Litera and Kira deal

Pages that carry this correction (15)

Claims about Casely, LlamaLab, Qualia, Clearbrief, Supio and EvenUp corrected or attributed to each vendor's own pages, first pass

· 8 pages

What changed

  • Removed the statement that Casely tracks environmental regulations and a monthly price estimate for it. Casely describes itself as legal AI for automatic document drafting, with a Starter plan at $98 and a Professional plan at $350 per month and a waitlist, not as a regulatory tracker. (1 page)
  • Replaced a description of LlamaLab's analysis features that its site does not make. LlamaLab sells medical record retrieval and AI analysis to plaintiff and mass-tort firms and lists HIPAA compliance on its pricing page. (1 page)
  • Replaced a statement that Qualia and Doma connect to title plants and underwriter databases. Qualia sells title and escrow production software and an agentic AI system, Qualia Clear, with pricing on request; we did not verify title plant connections and Doma's site is not covered here. (1 page)
  • Corrected Clearbrief's pricing. Its pricing page lists the Solo plan at $300 per user per month and custom pricing for Enterprise Unlimited; an earlier version said it publishes no pricing and gave an estimate we could not source. (1 page)
  • Reworded the American Arbitration Association references. Clearbrief lists the AAA among its adopters; we found no AAA rule or endorsement, so the page no longer calls it a partnership or says using Clearbrief shows compliance. Also removed an unsourced claim about how the tool works internally and a superlative. (2 pages)
  • Corrected the Supio profile and the Supio vs EvenUp comparison. Supio's site lists Supio Agent, Supio Intake and Case Communications and publishes no price (its pricing address redirects to the demo request page); product names, a pricing model, an accuracy figure and a call statistic from an earlier version are not on the site today and were removed. (2 pages)
  • Corrected the EvenUp profile. EvenUp publishes no price and offers a sales call; its leadership position, case-volume thresholds, hours saved per case and evaluation timeline had no source and were removed or attributed as vendor claims. (2 pages)

Source

Casely, home and pricing section (read October 25, 2026); LlamaLab, home and pricing page (read October 25, 2026); Qualia, home and pricing page (read October 25, 2026); Clearbrief, home and pricing section (read October 25, 2026); Supio, home and demo request page (read October 25, 2026); EvenUp, home page (read October 25, 2026)

Pages that carry this correction (8)

UK and EU AI rules: wording checked against the Bar Council guide and the EU AI Act text

· 3 pages

What changed

  • These pages summarised the Bar Council guide and the EU AI Act scope rules from secondary material or left them unread. We now cite the guide's current version (paragraphs 24 to 37: verification, confidentiality, no court-disclosure duty yet) and the Act's Article 2(1), Article 26, Article 99 and Annex III 8(a). The EU AI Act does not class legal AI as high-risk across the board, the EUR 35 million or 7% tier is for prohibited practices only, and the SRA did publish a warning notice on August 17, 2026. (2 pages)
  • Removed a count of federal courts with AI standing orders that had no source. Individual federal judges have issued such orders; we have not counted them. (2 pages)

Source

Bar Council, Considerations when using ChatGPT and generative AI software based on large language models (last reviewed November 25, 2025; read October 25, 2026); Regulation (EU) 2024/1689, Official Journal text: Articles 2, 26, 99 and Annex III (read October 25, 2026)

Pages that carry this correction (3)

Suno/Udio and Concord pages: court, judge, label status and figures reread against the record

· 2 pages

What changed

  • Corrected which labels are still suing and who the judge is. Warner dismissed its claims against Suno and Udio and UMG dismissed against Udio; Sony continues against Udio, and UMG and Sony against Suno, before Judge F. Dennis Saylor IV (Dkt. 253, July 2, 2026). Removed a summary judgment motion and hearing date, a trial date, a royalty rate and other settlement terms, two class-action captions with training-data figures, and a timeline graphic, none of which we could source. (1 page)
  • Corrected the court and the figures. Concord II was filed January 28, 2026 in the Northern District of California, not M.D. Tenn.; press lists 714 works in one exhibit and 20,517 in another, and we no longer say all 20,517 were torrented. Replaced the funding figure with Anthropic's own Series H announcement ($65 billion at a $965 billion post-money valuation) and removed compute-spend estimates, investor-pressure reports, RIAA amicus claims and unsourced details about Concord I. We have not read the Concord complaints (CourtListener returned 403 and rate limits), so Concord II facts are attributed to press and a case tracker. (1 page)

Source

Suno notice of supplemental authority, Dkt. 253, D. Mass. 1:24-cv-11611 (July 2, 2026); Anthropic, Series H announcement (May 28, 2026); AI Lawsuit Tracker, UMG, Sony & Warner v. Suno and Udio; AI Lawsuit Tracker, UMG/Concord/ABKCO v. Anthropic II

Pages that carry this correction (2)

Proposed Rule 707: status rewritten from the committee documents; EU AI Act December 2026 prohibitions added

· 1 page

What changed

  • The Rule 707 page said the Judicial Conference approved the rule in June 2025, cited over 3,400 comments and gave a December 2027 effective date. The committee documents do not support that. The rule has not been approved, the Advisory Committee held it back in May 2026 to study a revised draft (59 written comments by the February 16, 2026 deadline), the Standing Committee's September 2026 report lists it only as an information item, and there is no effective date. The page was rewritten from those documents. (1 page)

Source

Standing Committee, report to the Judicial Conference, September 2026 (revised August 27, 2026; read October 25, 2026); Advisory Committee on Evidence Rules, minutes of May 7, 2026 (read October 25, 2026); Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal text: Article 113 as amended (read October 25, 2026)

Pages that carry this correction (1)

Hawaii General Order 23-1 dated November 14, 2023 on the order itself

· 1 page

What changed

  • The summary line on Judge Otake's page gave November 13, 2023 as the issue date of General Order 23-1, repeating her procedures. The order itself is dated November 14, 2023 and stamped filed that day, so the page now gives November 14 and notes that her procedures say November 13. (1 page)

Source

D. Haw., General Order 23-1, In re Use of Unverified Sources (dated and filed November 14, 2023; read October 25, 2026); Judge Otake, General Civil Case Procedures (effective January 5, 2026; say the order was issued November 13, 2023)

Pages that carry this correction (1)

Canada, Australia, Singapore and India pages: wording checked against primary sources

· 7 pages

What changed

  • These pages stated several points without a primary source or against one. Singapore's guidance is non-binding and carries no procedural force. FASTER was a Supreme Court of India order of July 16, 2021, not a Delhi High Court measure. India's DPDP Rules were notified on November 13, 2025. SUPACE was launched under Chief Justice Bobde in April 2021. We removed an unsourced pending-case count and a National Judicial Data Grid figure, and a claim that the Spellbook and Canadian Bar Association deal was the first of its kind. (4 pages)
  • Goh Chin Cheng was decided on January 27, 2026 with grounds published May 4, 2026. We removed an unsourced count of US federal districts, a description of Australia's regulators that the joint statement does not support, a claim about how Spellbook's models were trained, and two unsourced national comparisons. (6 pages)

Source

Queensland Supreme Court Practice Direction 5 of 2025 (read October 25, 2026); Singapore MinLaw, Guide for using Generative AI in the Legal Sector (March 6, 2026; read October 25, 2026); Goh Chin Cheng v Choco Up SG Pte Ltd [2026] SGHCR 13 (read October 25, 2026); PIB, Digital Personal Data Protection Rules notified November 13, 2025 (read October 25, 2026); Supreme Court of India order of July 16, 2021 on FASTER (read October 25, 2026)

Pages that carry this correction (7)

Funding figures corrected against company announcements on ten more pages

· 10 pages

What changed

  • Corrected the Blackstone and Norm Law timing. Blackstone's $50M went into Norm Ai in November 2025 (Norm Law launched November 20, 2025), about six weeks after Spellbook's October 9, 2025 Series B, not at the same time and not into Norm Law directly. (2 pages)
  • Corrected Spellbook. The Series B was announced October 9, 2025, not April 2026; the more than $80M raised to date is BetaKit's figure, and Spellbook's announcement does not call Threshold Ventures a new investor. (3 pages)
  • Corrected Harvey. Its announcements give dated rounds and no cumulative total, harvey.ai lists 700k+ daily agent tasks, and no source read supports Microsoft investing in its Series C. (4 pages)
  • Corrected Crosby, Lawhive and Darrow. Crosby's $60M Series B is dated March 31, 2026 and its $85.8M total is AlleyWatch's figure; Lawhive's $35M is annual revenue (a company figure reported by Fortune), not ARR; Darrow announced a $35M Series B and no source gives a $44M total. (3 pages)

Source

Spellbook, Series B announcement (October 9, 2025); BetaKit, Spellbook Series B: $350M post-money, more than $80M raised to date (October 9, 2025); LawFuel, The Blackstone Theory of Legal Services (April 28, 2026): Norm Law launched November 20, 2025; Harvey, $550M at a $15.5B valuation (September 9, 2026); Harvey, agents by the numbers (read October 25, 2026); Crosby, Series B and Planting our Flag (March 31, 2026); AlleyWatch daily funding report (March 31, 2026); Darrow, $35M Series B coverage (September 2023)

Pages that carry this correction (10)

Legal AI funding and valuation figures reread against company announcements and filings

· 7 pages

What changed

  • Removed a PitchBook-attributed legal AI funding total with a deal count, a funding-to-revenue ratio and a 2019-2023 comparison that we could not source. PitchBook, as reported by Axios on November 14, 2025, counted about $5 billion of legal tech venture funding through November 2025 and $2.8 billion in all of 2024. (2 pages)
  • Corrected Thomson Reuters. Its Legal Professionals segment had $2,868 million in revenue in 2025 (not $6.8 billion), and the annual report states more than $200 million invested in AI in 2024; the $650M figure is the 2023 Casetext acquisition price. (1 page)
  • Harvey's investors now follow its September 9, 2026 announcement (Sequoia, Kleiner Perkins, a16z, GIC and others; Diffusion and Lightspeed co-led) and its March 25, 2026 post (GIC and Sequoia co-led). The total above $1 billion is attributed to press coverage because Harvey's own posts give no total. Removed a market-size figure and an unsourced reading of OpenAI's role. (2 pages)
  • Removed Spellbook trailing revenue and ARR estimates and the multiples built on them: no source gives them, and BetaKit reports the CEO gave no figure. The $350M valuation is BetaKit's figure, 3.5x the CEO's $100M ARR target. (3 pages)
  • Corrected Spellbook's timeline. Spellbook Library launched in July 2025, the Series B was October 9, 2025 and the CBA exclusive followed on March 3, 2026. Inovia led the earlier round; Threshold Ventures took part in the Series B. Blackstone's $50M went to Norm Ai in November 2025, weeks after the Series B, not the same week. (3 pages)
  • Corrected Supio's funding claims: it raised its $60 million Series B on April 30, 2025 ($91 million in total), and EvenUp reported $385 million in total capital in October 2025, so Supio is not the best-funded personal-injury platform. (1 page)
  • Removed the claim that $5B+ valuations sit on under $100M ARR (Legora reported ARR above $200 million at its October 2026 Series E) and fixed a corrupted valuation figure in an FAQ answer. (1 page)

Source

Harvey, $550M at a $15.5B valuation (September 9, 2026); Harvey, $200M at an $11B valuation (March 25, 2026); Spellbook, Series B announcement (October 9, 2025); BetaKit, Spellbook Series B (October 9, 2025) and $100M ARR claim (March 4, 2026); Legora, Series E announcement (October 21, 2026); Supio, Series B announcement (April 30, 2025); EvenUp, Series E announcement (October 7, 2025); Thomson Reuters, Annual Report 2025, SEC Form 40-F

Pages that carry this correction (7)

Crosby, Norm Ai, Luminance, Harvey, Lawhive and Atrium funding figures corrected against primaries

· 14 pages

What changed

  • Removed Crosby's $85.8M as a company figure. Crosby's own posts give a $60M Series B (March 31, 2026) and a $20M Series A (October 8, 2025) and state no total; the total is AlleyWatch's, so it appears only attributed to AlleyWatch. (3 pages)
  • Made one attributed version of the Norm story: Blackstone's new $50M investment went into Norm Ai, the company, and Norm Law is its affiliated AI-native law firm, launched on November 20, 2025 (Reuters, as listed in Norm Ai's newsroom). (7 pages)
  • Replaced a Luminance link that returned 404 with the CEO's post on its $75M Series C, which gives total funding of over $115M and no valuation or exact announcement day. (1 page)
  • Corrected Lawhive and Atrium. Lawhive's $35M is annual revenue (a CEO figure reported by Fortune), not money raised, and the unsourced 70% and 50% claims were removed; Atrium raised $75.5M and shut down its operations in March 2020 while its separate law firm kept operating (TechCrunch). (1 page)
  • Fixed the chronology: Sequoia published its essay on March 5, 2026, not on a Monday night, and the Crosby funding detail on this page was added after the March 31 announcement, so it is now labelled as an update. (1 page)
  • Harvey: the $11B valuation belongs to the March 2026 round, not the Series D ($300M at $3B, February 2025); the Series C ($100M at $1.5B, July 2024) and Series D now cite Harvey's own announcements. (2 pages)

Source

Crosby, Series B and Planting our Flag (March 31, 2026); Crosby, Series A and Smarter Agreements, Faster (October 8, 2025); Norm Ai, $120M Series C announcement (July 7, 2026), which describes Norm Law, LLP as an affiliated law firm; Norm Ai newsroom, listing Reuters (November 20, 2025), on the new $50M Blackstone investment in Norm Ai and the launch of Norm Law; Luminance, CEO post on the Series C (February 27, 2025); Harvey, Series C announcement (July 23, 2024); Harvey, Series D announcement (February 12, 2025); Sequoia Capital, Services: The New Software (March 5, 2026); Fortune, Lawhive Series B (February 5, 2026); TechCrunch, Atrium shuts down (March 3, 2020)

Pages that carry this correction (14)

Unsourced legal AI funding total retired from the remaining pages

· 10 pages

What changed

  • Removed a legal AI funding total that no source supported, including from a section heading, a page title and link text. Where a funding figure remains, it is the PitchBook total reported by Axios on November 14, 2025: about $5 billion of legal tech venture funding through November 2025. (10 pages)

Source

Axios Pro, Legal-tech's venture valuation escalation (November 14, 2025, citing PitchBook)

Pages that carry this correction (10)

Florida state courts: no judge-by-judge or circuit AI disclosure orders since June 15, 2026

· 12 pages

What changed

  • These pages said the Florida Supreme Court had issued no statewide AI order and that individual judges or circuits (Miami-Dade, Broward and others) set AI disclosure rules for state court filings. That is outdated: the Court's order AOSC26-12 (May 28, 2026, effective June 15, 2026) says courts may not impose their own AI disclosure or certification requirements, and amended Rule 2.515(d)(2) has each signer represent that the cited authorities exist and are accurately cited (read October 25, 2026). The pages now say so. Florida's federal judges keep their own orders. (12 pages)

Source

Florida Supreme Court, AOSC26-12, May 28, 2026 (flcourts-media.flcourts.gov, read 2026-10-25); The Florida Bar News, May 29, 2026

Pages that carry this correction (12)

EU AI Act law-firm guide: the 2 December 2026 date for the new Article 5 prohibitions added

· 1 page

What changed

  • The page did not give the date on which the new Article 5 prohibitions apply. It now states 2 December 2026 (Article 5(1), first subparagraph, points (ba) and (bb), and Article 5(1a) and (1b)), taken from the Official Journal text of Regulation (EU) 2026/1744 and Article 113 as amended. (1 page)

Source

Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal text: Article 113 as amended (op.europa.eu, read October 29, 2026)

Pages that carry this correction (1)

AI ethics opinions 2026: California proposal status and source links

· 1 page

What changed

  • The California row said only that the comment period on the proposed AI comments to six conduct rules closed May 4, 2026. The State Bar's page shows COPRAC revised the proposals and approved a second round on June 12, 2026, with a comment deadline of August 6, 2026. The row now says so; we still found no final version. (1 page)
  • Fixed a stray line-break tag that printed as text in the source cells of the table. (1 page)

Source

State Bar of California, public comment page on the proposed AI amendments to the Rules of Professional Conduct (read October 25, 2026)

Pages that carry this correction (1)

Unnamed "Legal AI Summit (multiple cities)" replaced by the dated event we could verify

· 1 page

What changed

  • Removed a mention of a "Legal AI Summit" in multiple cities that we could not tie to an organizer page. The guide now names the Legal 500 AI Summit in London (November 18, 2026, a save the date that can change) and the AI and the Law Conference in Minneapolis (November 9-10, 2026), each with the date we read on the organizer's page. (1 page)

Source

Legal 500 Events, AI Summit 2026 page (read October 25, 2026); Minnesota CLE, 2026 AI and the Law Conference (read October 25, 2026)

Pages that carry this correction (1)

CoCounsel: Thomson Reuters publishes starting prices, so 'quote-only' was wrong

· 23 pages

What changed

  • These pages said CoCounsel could only be priced through a sales quotation. Thomson Reuters' own page lists starting prices for in-house teams: CoCounsel Essentials from $340.85 per month, Practical Law Dynamic Tool Set with CoCounsel Essentials from $882.30 per month and CoCounsel Legal from $1,694.90 per month (read October 25, 2026). Law-firm plans are still quoted through a configurator, and the pages now say so. (23 pages)

Source

Thomson Reuters, CoCounsel plans for in-house legal teams (read 2026-10-25)

Pages that carry this correction (23)

CoCounsel: pages that mixed 'quote-only' with the price Thomson Reuters publishes

· 57 pages

What changed

  • These pages said CoCounsel as a whole is quote-only or has no published price, or gave third-party tier figures without naming the source. Thomson Reuters lists CoCounsel from $340.85 per month for in-house teams (read October 25, 2026); law-firm plans are quoted, and third-party tier figures are now labeled as estimates Thomson Reuters has not confirmed. (57 pages)

Source

Thomson Reuters, CoCounsel plans for in-house legal teams (read 2026-10-25)

Pages that carry this correction (57)

Unsourced claims about Clio's SOC 2 status replaced with a note to ask Clio (third pass: Clio comparison, use-case and guide pages)

· 9 pages

What changed

  • Replaced unsourced statements about Clio's SOC 2 status, and about where practice data stays, with a note to ask Clio for its current SOC 2 report and AI data terms. Clio's own site was not available for checking on October 25, 2026, so no Clio security page was read and the audit status is not stated as fact. (9 pages)

Source

Clio, security page (answered 403 on October 25, 2026; no statement taken from it, listed as where to ask)

Pages that carry this correction (9)

Clearbrief awards and AAA-ICDR wording checked against the organizer, Clearbrief's press page and LawNext; EvenUp 'market leader' link text replaced

· 8 pages

What changed

  • Named the organizer, year and categories of Clearbrief's awards. The Legalweek Leaders in Tech Law Awards 2026 (Law.com Events) list Clearbrief as winner in Litigation Technology and Technology for Good, and Clearbrief announced a 2025 LegalTech Breakthrough Award for Legal Compliance Solution of the Year. Earlier text said only that the vendor said it won. (2 pages)
  • Corrected our October 25 wording on the American Arbitration Association. Clearbrief's press page lists a partnership with AAA-ICDR and LawNext reported on January 14, 2025 that AAA-ICDR would offer Clearbrief to its arbitrators and mediators; we found no AAA announcement and no AAA rule naming Clearbrief, so the page says so. (2 pages)
  • Replaced related-reading link text that stated 'EvenUp is the market leader' as our own finding. EvenUp calls itself the #1 proactive AI platform for PI, which is a vendor claim. (5 pages)

Source

Law.com Events, Legalweek Leaders in Tech Law Awards 2026 winners and finalists (read October 25, 2026); Clearbrief, press page (read October 25, 2026); Clearbrief, LegalTech Breakthrough Award announcement (November 13, 2025); LawNext, AAA-ICDR and Clearbrief (January 14, 2025)

Pages that carry this correction (8)

Bartz v. Anthropic page rewritten from the court record: settlement terms without a source removed

· 1 page

What changed

  • Removed settlement terms that the Settlement Agreement does not contain: a $500 million Year 1 and $250 million a year schedule over five years (the Agreement sets four installments over about two years), an obligation to license all training data, a public registry of training sources, a 'do not train' registry, Special Master or third-party audits of Anthropic's pipeline, and a $300 million author sub-fund. Also removed quoted Slack messages and executive emails, Z-Library, and a 500,000-work library size, none of which appear in the court's orders. (1 page)
  • Corrected the date and the subject. The settlement received final approval on July 20, 2026, not in March 2026, and the case is about copying books, not liability for AI output; the court made no finding on Section 230, terms of service or product liability. The title and description now say so. (1 page)

Source

Order on fair use, Dkt. 231 (June 23, 2025); Class Action Settlement Agreement, Dkt. 362-3; Order granting final approval and judgment, Dkt. 680 (July 20, 2026)

Pages that carry this correction (1)

Bartz v. Anthropic mentions on other pages: wrong court and holding removed

· 3 pages

What changed

  • Removed a 2024 appeals-court decision about audio analysis that the page attributed to the name Bartz, and the statement that Suno relies on it. We found no such case; the only Bartz case we cite, Bartz v. Anthropic, is a 2025 order of the Northern District of California about books, and we have not read Suno's briefing. (1 page)
  • Corrected what the Bartz order held. Judge Alsup held that training on the books at issue was fair use and that keeping pirated copies in a central library was not excused by fair use (Dkt. 231); he did not hold that pirated training data is categorically outside fair use, and the order concerns books, not music. Removed a claim about how Concord's lawyers use Bartz that we had not read, and a combined exposure figure that added the settled Bartz case to Concord's claim. (1 page)
  • Corrected the judge. Judge Alsup wrote the June 23, 2025 fair use order; the July 20, 2026 final approval was signed by Judge Araceli Martinez-Olguin (Dkt. 680). (1 page)

Source

Order on fair use, Dkt. 231 (June 23, 2025); Order granting final approval and judgment, Dkt. 680 (July 20, 2026)

Pages that carry this correction (3)

Judge Starr's order no longer called 'one of the nation's first' or 'the model for dozens of later orders'

· 2 pages

What changed

  • Pages called Judge Starr's May 2023 AI order 'one of the nation's first mandatory AI certification orders' or 'the model for dozens of later orders across the country'. Neither claim has a source. The pages now say the order is widely reported as the first from a federal judge, that other courts and judges issued their own AI orders afterward, and point to the court page for what the order required. (2 pages)

Source

Judge Starr, chambers page (txnd.uscourts.gov, read October 24, 2026); Judge Starr, Civil Rules page (txnd.uscourts.gov, read October 24, 2026)

Pages that carry this correction (2)

Starr 'first federal AI standing order': attributed on 12 pages

· 10 pages

What changed

  • Pages that called Judge Starr's May 2023 order the first federal AI standing order now attribute that to press reports, with the court page as the source for what the order required. (10 pages)

Source

Judge Starr, chambers page (txnd.uscourts.gov, read October 24, 2026); Judge Starr, Civil Rules page (txnd.uscourts.gov, read October 24, 2026)

Pages that carry this correction (10)

Mata v. Avianca: Starr's order no longer called 'the first such order'; attributed to press reports

· 1 page

What changed

  • The page said Judge Starr issued 'the first such order' just before the Mata sanctions. It now says he issued a May 2023 AI standing order that press coverage reports as the first from a federal judge, with the court page as the source for what the order required. (1 page)

Source

Judge Starr, chambers page (txnd.uscourts.gov, read October 24, 2026); Judge Starr, Civil Rules page (txnd.uscourts.gov, read October 24, 2026)

Pages that carry this correction (1)

What courts require: Northern District of Illinois sentence

· 1 page

What changed

  • Removed a sentence saying Magistrate Judge Fuentes adopted a May 2023 order requiring disclosure of the AI tool, which rested on a secondary source. His civil order, read October 24, 2026, points counsel to the Illinois Supreme Court AI policy and Rule 11 and has no disclosure requirement. (1 page)

Source

Judge Fuentes, civil standing order, section Artificial Intelligence (ilnd.uscourts.gov, read October 24, 2026); Judge Coleman, standing order page (ilnd.uscourts.gov, read October 24, 2026)

Pages that carry this correction (1)

Judges Woods and Chen: no AI standing order in the primary orders

· 2 pages

What changed

  • Removed the title and description that promised an AI standing order with disclosure, filing-format and sanctions rules for Judges Woods and Chen. Their orders on the court sites do not mention AI (read October 23, 2026); the FAQ now says so and links the orders. (2 pages)

Source

Judge Woods, Individual Rules of Practice in Civil Cases, revised March 19, 2025 (nysd.uscourts.gov); Judge Chen, standing orders and guidelines (cand.uscourts.gov)

Pages that carry this correction (2)

Judge Starr: 'first' claim on three judge pages

· 3 pages

What changed

  • The chambers page gives no date and does not call the order the first. The pages now attribute 'first' to press coverage instead of stating it as fact. (3 pages)

Source

Judge Starr, chambers page (txnd.uscourts.gov, read October 24, 2026)

Pages that carry this correction (3)

Judge Newman: effective date of the AI provision

· 1 page

What changed

  • Replaced July 14, 2023 with the date in the order itself: its notice reads "New AI Provision Effective as of August 27, 2025". (1 page)

Source

Judge Newman, standing civil order of 8/27/25 (ohsd.uscourts.gov, read October 24, 2026)

Pages that carry this correction (1)

Judge Fuentes and neighbouring Illinois pages: order wrongly described

· 4 pages

What changed

  • Removed claims that Judge Fuentes first required disclosure of AI use and later withdrew or softened the rule. His civil order has no disclosure requirement and no earlier order was found on a court site. (4 pages)

Source

Judge Fuentes, civil standing order, section Artificial Intelligence (ilnd.uscourts.gov, read October 24, 2026); Judge Coleman, standing order page (ilnd.uscourts.gov, read October 24, 2026)

Pages that carry this correction (4)

Judge Cole: terms of an order that cannot be read

· 3 pages

What changed

  • Removed the stated terms of an AI order attributed to Magistrate Judge Cole. The court's link to that order returned 404 on October 24, 2026 and he is not on the judges list, so the terms are no longer restated. (3 pages)

Source

Judge Coleman, standing order page (ilnd.uscourts.gov, read October 24, 2026); Judge Fuentes, civil standing order, section Artificial Intelligence (ilnd.uscourts.gov, read October 24, 2026)

Pages that carry this correction (3)

Judge Baylson: 'one of the earliest' claim

· 1 page

What changed

  • Replaced the ranking claim with what the order itself shows: it is dated June 6, 2023. Also fixed the link to the order, which had gone to a 404. (1 page)

Source

Judge Baylson, Standing Order re Artificial Intelligence, dated 6/6/2023 (paed.uscourts.gov, read October 24, 2026)

Pages that carry this correction (1)

Harvey and CoCounsel: agent and accuracy figures without a source

· 12 pages

What changed

  • Removed a figure of 25,000 custom agents and a CoCounsel accuracy figure that neither vendor's page publishes. Harvey's task and customer counts now follow harvey.ai (700,000+ agent tasks, 3,500+ customers, 200,000+ professionals, checked October 23, 2026). (12 pages)

Source

Harvey (harvey.ai); Thomson Reuters, CoCounsel Legal (legal.thomsonreuters.com)

Pages that carry this correction (12)

Copilot pricing: expired promotion and unreadable base prices

· 1 page

What changed

  • Removed an expired June 30, 2026 promotion and combined cost totals built on Microsoft 365 base-license prices that could not be re-read, and pointed to Microsoft's current plans. (1 page)

Source

Microsoft 365 Copilot, pricing (microsoft.com)

Pages that carry this correction (1)

Claude API price: current model and rate

· 1 page

What changed

  • Replaced the Opus 4.7 rate ($5 and $25 per million tokens) with the current Opus 5.5 rate on Anthropic's pricing page ($4 input, $20 output, read October 24, 2026). (1 page)

Source

Anthropic, API pricing (platform.claude.com); Anthropic, Claude plans and pricing (claude.com)

Pages that carry this correction (1)

Andersen v. Stability AI: stale procedural status

· 1 page

What changed

  • Replaced April 2026 text that said the case was heading toward discovery with the docket's procedural status as of October 24, 2026, with ECF numbers. (1 page)

Source

N.D. Cal., Andersen v. Stability AI, No. 3:23-cv-00201, ECF 778 (CourtListener RECAP copy); N.D. Cal., Andersen v. Stability AI, ECF 623 (CourtListener RECAP copy)

Pages that carry this correction (1)

Visalaw.ai: product claims re-read against the vendor

· 1 page

What changed

  • Removed 'GEN Core' and '200,000+ sources' claims that Visalaw.ai's and AILA's pages do not carry. The page now states public pricing from $220 per month, SOC 2 Type II and the Research Agent in release 2.1. (1 page)

Source

Visalaw.ai (visalaw.ai); American Immigration Lawyers Association (aila.org)

Pages that carry this correction (1)

Deepfakes guide: TAKE IT DOWN Act facts and state counts

· 1 page

What changed

  • Corrected the TAKE IT DOWN Act facts: it was signed May 19, 2025 (Public Law 119-12), the maximum for offenses involving minors is 3 years, not 10, and the Act does not give victims a private lawsuit. (1 page)
  • Removed the state counts (43 states overall, with counts per category) because we could not tie them to a primary source. (1 page)

Source

TAKE IT DOWN Act, Public Law 119-12 (govinfo.gov)

Pages that carry this correction (1)

Third Circuit hub: ROSS appeal already decided

· 1 page

What changed

  • The Third Circuit hub described Thomson Reuters v. ROSS as on appeal. The court affirmed on September 29, 2026 (No. 25-2153); the hub now says so. (1 page)

Source

Third Circuit, Thomson Reuters v. ROSS Intelligence, No. 25-2153, opinion of September 29, 2026 (ca3.uscourts.gov)

Pages that carry this correction (1)

ROSS page: appeal status out of date

· 1 page

What changed

  • Replaced the appeal-pending status on the ROSS page with the Third Circuit's September 29, 2026 affirmance (No. 25-2153), as a dated section with matching FAQ and structured data. (1 page)

Source

Third Circuit, Thomson Reuters v. ROSS Intelligence, No. 25-2153, opinion of September 29, 2026 (ca3.uscourts.gov)

Pages that carry this correction (1)

Robin AI: profile replaced by a dated status page

· 5 pages

What changed

  • Removed the profile that presented Robin AI as an active product and the links to it from the AI tools hub and three related guides. The page now says what the sources report (a distressed-sale listing in October 2025, its managed services moving to Scissero in December 2025, Microsoft hiring former staff in January 2026) and what is not confirmed. (5 pages)

Source

Legal IT Insider, Robin AI listed for distressed sale (October 28, 2025); Artificial Lawyer, the Scissero and Robin AI deal (December 10, 2025); Bloomberg Tax, Microsoft brings former Robin AI employees into the fold (January 2026)

Pages that carry this correction (5)

Best AI for estate planning: unsourced time-saved figures, WealthCounsel price and Trust & Will AI claim removed; Gavel plans taken from its pricing page

· 1 page

What changed

  • Removed time-saved figures we could not source (this page said '80% of the document stays the same', '3-4 hours of drafting per plan', '4-6 hours down to 1-2' and '10-15 hours per week'). The page now says we have no measured figure and tells readers to time a few of their own plans. (1 page)
  • Removed WealthCounsel's price of $150-200 a month, the claim of built-in state-specific compliance checks and the claim of thousands of contributing attorneys. None of them was on a page we could read. (1 page)
  • Removed the claim that Trust & Will uses AI-assisted intake and document generation; the page now describes the plans its own site lists. (1 page)
  • Replaced 'starting at $83 a month' for Gavel with the plan list on its pricing page, with the date it was read. (1 page)

Source

Gavel, pricing (Lite, Standard, Pro and Scale plans; read 2026-10-23); AI Vortex sourcing rule (a figure or claim on a page needs a published source; none found for the removed claim)

Pages that carry this correction (1)

Montana, Nebraska, Louisiana: unverified Missouri opinion and a wrong New York opinion number removed

· 4 pages

What changed

  • The Montana and Nebraska pages cited Missouri Informal Opinion 2024-11 as guidance. We could only find it described in a secondary source, not on an issuing body's site, so those references were removed and the pages point to Kentucky Ethics Opinion E-457, which we read in full. (3 pages)
  • The Louisiana page named a numbered formal opinion as New York's AI opinion. New York's document is the NYSBA Task Force report of April 2024, not a numbered opinion, so the page now says that. (1 page)

Source

Kentucky Bar Association, Ethics Opinion KBA E-457, issued March 15, 2024 (kybar.org, read October 29, 2026); NYSBA Task Force on Artificial Intelligence, Report and Recommendations, April 2024 (nysba.org, read October 29, 2026); Louisiana Supreme Court, letter on artificial intelligence, January 22, 2024 (lsba.org, read October 29, 2026)

Pages that carry this correction (4)

State ethics opinions: wrong or unverifiable numbers

· 8 pages

What changed

  • Replaced opinion numbers that do not exist in the bars' own indexes with the audited ones (Oregon 2025-205, Pennsylvania Joint Formal Opinion 2024-200, North Carolina 2024 FEO 1, Washington Advisory Opinion 2025-05) or removed them (California, New York City, Wisconsin, Virginia). (8 pages)

Source

Oregon State Bar, Formal Opinion 2025-205 (osbar.org); Washington State Bar Association, Advisory Opinion 2025-05 (ao.wsba.org); Virginia Supreme Court, LEO 1901 (vacourts.gov); North Carolina State Bar, 2024 Formal Ethics Opinion 1 (ncbar.gov); Pennsylvania Bar Association and Philadelphia Bar Association, Joint Formal Opinion 2024-200 (philadelphiabar.org)

Pages that carry this correction (8)

Wording about our reading process removed

· 107 pages

What changed

  • Rewrote wording that described our own research process. Sections that used to carry a note about our own checking now read "Before you rely on this" or "Open questions" and speak to the reader. (107 pages)

Source

AI Vortex editorial rule of October 22, 2026 (a page states what the sources say and does not narrate our research process)

Pages that carry this correction (107)

Oregon sanction figure removed

· 112 pages

What changed

  • Removed an Oregon sanction figure that did not match any court record. The Oregon case in our sanctions dataset is Couvrette v. Wisnovsky (D. Or.): a $15,500 sanction plus $94,700 in costs, $110,204 in total. (111 pages)
  • Removed an Oregon case name and attorney description that did not match any court record. (2 pages)

Source

Couvrette v. Wisnovsky, D. Or. (order PDF in the AI Hallucination Cases database); Damien Charlotin, AI Hallucination Cases database, the dataset behind our sanctions tracker (damiencharlotin.com)

Pages that carry this correction (112)

Oregon bar opinion: per-citation penalty formula

· 7 pages

What changed

  • Removed a dollar-penalty formula attributed to Oregon. Formal Opinion 2025-205 sets no dollar penalties. (7 pages)

Source

Oregon State Bar, Formal Opinion 2025-205 (osbar.org)

Pages that carry this correction (7)

Legora: customer and market counts out of date

· 5 pages

What changed

  • Replaced Legora's older scale figures with the ones in its own Series E announcement: more than 2,400 law firms and in-house legal teams across more than 80 markets. (5 pages)

Source

Legora, Series E announcement (legora.com, October 21, 2026)

Pages that carry this correction (5)

Case facts corrected: Morgan v. V2X and Brigandi

· 17 pages

What changed

  • Corrected Morgan v. V2X to D. Colo. No. 1:25-cv-01991, order of March 30, 2026. Courts, dates and requirements that did not match that order were removed. (16 pages)
  • Replaced an unsupported Brigandi sanction amount and date with the dataset's Couvrette v. Wisnovsky amounts: $15,500 plus $94,700 in costs, $110,204 in total. (1 page)

Source

Damien Charlotin, AI Hallucination Cases database, the dataset behind our sanctions tracker (damiencharlotin.com)

Pages that carry this correction (17)

Case facts corrected: Park v. Kim and other entries

· 24 pages

What changed

  • Corrected Park v. Kim to the Second Circuit opinion (No. 22-2057): the court referred the lawyer to its Grievance Panel and imposed no fine. (7 pages)
  • Removed three case entries that are not in the sanctions dataset. (4 pages)
  • Corrected the Cherry Hill, New Jersey sanction of April 27 to the Bunce case (E.D. Pa., $5,000). (14 pages)
  • Replaced Q1 2026 sanction totals that came from a secondary source with figures computed from the sanctions dataset. (3 pages)

Source

Park v. Kim, 2d Cir., No. 22-2057, opinion of January 30, 2024 (govinfo.gov); Bunce v. Visual Technology Innovations, E.D. Pa. (ruling reported April 2026); Damien Charlotin, AI Hallucination Cases database, the dataset behind our sanctions tracker (damiencharlotin.com)

Pages that carry this correction (24)

Law firm AI committee guide: unsourced February 2025 'ABA Commission' guidance removed; cites ABA Formal Opinion 512 (July 29, 2024)

· 2 pages

What changed

  • The page said an ABA 'Commission' released formal guidance in February 2025 establishing attorney obligations for AI use. We could not tie it to any ABA source, so it was removed; the page now cites ABA Formal Opinion 512 (July 29, 2024), which applied the existing Model Rules to lawyers' use of generative AI. (2 pages)

Source

American Bar Association, Formal Opinion 512, Generative Artificial Intelligence Tools (July 29, 2024)

Pages that carry this correction (2)

vLex, Westlaw and Lexis: price figures and superlatives removed

· 3 pages

What changed

  • Removed price figures, a 'free through state bars' claim for vLex and unsourced rankings of Westlaw, Lexis and vLex. The pages now say vLex shows no public price and Lexis sells by quote; the Westlaw wording was corrected again on October 29, 2026, since Thomson Reuters lists Westlaw Advantage pricing online only for firms of up to 10 attorneys. (3 pages)

Source

vLex, Vincent AI (vlex.com); Thomson Reuters, Westlaw (legal.thomsonreuters.com); LexisNexis, Protégé (lexisnexis.com; sold by quote)

Pages that carry this correction (3)

Spellbook and Clio Duo: unsourced prices removed

· 50 pages

What changed

  • Removed Spellbook prices and a Clio Duo price that the vendors do not publish in that form. Spellbook's pricing page says pricing is custom. (50 pages)

Source

Spellbook, pricing page (spellbook.legal)

Pages that carry this correction (50)

Harvey and CoCounsel: estimated per-seat prices removed

· 57 pages

What changed

  • Removed per-seat prices for Harvey and CoCounsel that came from market estimates. Harvey publishes no price. Thomson Reuters quotes CoCounsel firm plans; updated October 25, 2026: it also lists starting prices for in-house teams (see the October 25 corrections). (57 pages)

Source

Harvey (harvey.ai; no public price); Thomson Reuters, CoCounsel Legal (legal.thomsonreuters.com; firm plans quoted)

Pages that carry this correction (57)

Judge Starr: AI certification is no longer required

· 15 pages

What changed

  • Corrected the Judge Starr entry: a standing AI certification is no longer required. The current rules are N.D. Tex. LR 7.2(f) and LCrR 47.2(e), editions in force September 1, 2026. (15 pages)

Source

Judge Brantley Starr, N.D. Tex. (judge page); N.D. Tex. local civil rules (LR 7.2(f)) and criminal rules (LCrR 47.2(e))

Pages that carry this correction (15)

Spellbook: company announcement date corrected

· 6 pages

What changed

  • Corrected the date of Spellbook's $50M Series B: BetaKit reported it on October 9, 2025, not April 2026. (3 pages)
  • Removed predictions about how Spellbook's pricing terms will change in 2026 to 2028; no public source supports them. (5 pages)

Source

BetaKit, Spellbook raises $50M USD Series B led by Khosla Ventures (published October 9, 2025)

Pages that carry this correction (6)

Harvey: stale company figures replaced

· 4 pages

What changed

  • Updated Harvey's stated valuation from its 2025 figure to the $15.5 billion in Harvey's own announcement of its latest round. (4 pages)

Source

Harvey, Harvey raises $550M at a $15.5B valuation (harvey.ai)

Pages that carry this correction (4)

A&O Shearman agentic playbook: unsourced Harvey price and the first-year totals built on it removed

· 1 page

What changed

  • Removed a per-user monthly price for Harvey and the annual and first-year totals built on it. Harvey publishes no price, so the page now says to get a quote, lists the cost lines to plan for and gives no total. (1 page)

Source

Harvey (harvey.ai; no public price)

Pages that carry this correction (1)

Best AI for family law no longer says AI will cut document prep time by 40%

· 1 page

What changed

  • Removed the claim that early adopters in family law 'will cut document prep time' by a stated percentage. We found no measured figure for family law, so the sentence no longer gives one; the wording now says adopters can cut prep time and tells readers to test it on their own matters. (1 page)

Source

AI Vortex sourcing rule (a figure or claim on a page needs a published source; none found for the removed claim)

Pages that carry this correction (1)

eImmigration $55 per user per month removed and Docketwise price corrected to its listed $80 to $140 range

· 7 pages

What changed

  • Removed the per-user monthly price we had given for eImmigration, with the claims built on it ('most affordable', a monthly total for five people, 'best value per dollar'). eImmigration's pricing page sends buyers to a demo request and shows no price; the page now says so and tells readers to get a quote. (5 pages)
  • Pages gave an outdated per-user monthly price for Docketwise. Docketwise's pricing page lists Basic, Pro and Advanced plans, each with its own price depending on plan and billing period; the pages now point to its listed prices with the date read. (7 pages)

Source

eImmigration, pricing page (routes to a demo request, no price shown; read 2026-10-20); Docketwise, pricing (Basic, Pro and Advanced plans, $80 to $140 per user per month depending on plan and billing period; read 2026-10-20)

Pages that carry this correction (7)

DISCO $20,000-100,000 a year range and the '$50K/year' label on Everlaw and DISCO removed; DISCO is quote only

· 2 pages

What changed

  • Removed an annual price range for DISCO, the minimum-spend reasoning and the break-even claims built on it. DISCO publishes no price and sells by quote, so the page now says that. (1 page)
  • The Logikcull page described Everlaw and DISCO as platforms with a fixed annual price. We found no published price for either, so the figure was removed. (1 page)

Source

DISCO, pricing (csdisco.com)

Pages that carry this correction (2)

DISCO corporate history corrected

· 3 pages

What changed

  • Removed incorrect statements about DISCO's corporate history. The company's current status is the one in its own SEC filings. (3 pages)

Source

CS Disco Q2 fiscal 2026 earnings release (SEC filing, August 5, 2026)

Pages that carry this correction (3)

Clearbrief and DISCO: prices replaced

· 3 pages

What changed

  • Replaced Clearbrief's price with the Solo plan on its pricing page ($300 per user per month) and recalculated the trial-lawyer stack total. DISCO's derived per-GB and per-matter figures were removed; it publishes no price. (3 pages)

Source

Clearbrief, pricing (clearbrief.com); DISCO, pricing (csdisco.com)

Pages that carry this correction (3)

Michigan: certification rule cited as MCR 1.109(E), not MCR 2.114

· 1 page

What changed

  • The page cited MCR 2.114 for the rule that a signature certifies a factual and legal basis for a filing. That certification is in MCR 1.109(E); a reader pointed out the old number. The page now cites MCR 1.109(E) and notes the former number. (1 page)

Source

Michigan Court Rules of 1985, Rule 1.109(E) (courts.michigan.gov, updated July 31, 2026, read October 29, 2026)

Pages that carry this correction (1)