The dated procedural map
| Document | Observed state | What it supports | What it does not establish |
|---|---|---|---|
| Thomson Reuters v. ROSS, Feb. 11, 2025 | District-court summary-judgment opinion | On that record, the court rejected ROSS's fair-use defenses concerning Westlaw headnotes used to build a competing legal-research system. | That every model, dataset or purpose has the same market effect. |
| Kadrey v. Meta, June 25, 2025 | Summary judgment for Meta on the named plaintiffs' training claim | The court ruled for Meta on the record presented by those plaintiffs. | A general rule that training on copyrighted works is always lawful; the order expressly describes its reach as limited. |
| Bartz v. Anthropic, July 20, 2026 | Final class-settlement approval | The court approved the settlement for the covered class. | A merits holding that all alleged acquisition or training conduct was lawful or unlawful. |
| Copyright Office Part 3, May 2025 pre-publication version | Policy and legal analysis | A federal expert agency's analysis of licensing, fair use and market questions. | A binding judicial judgment on a particular vendor or buyer. |
Do not promote allegations into holdings
A complaint tells the court what a party alleges. A motion asks for a result. An order explains what the court decided at that stage. A settlement can resolve claims without deciding their merits. Record the document type, court, date, claim, works, acquisition path, use, market evidence and procedural state before summarizing a headline.
This distinction matters in procurement. A vendor may describe a lawsuit as resolved while an appeal, separate claim or different class remains outside the document being cited. Conversely, a pending allegation is not proof of infringement. Preserve neutral language until the current docket supports more.
The fair-use comparison remains fact-specific
ROSS and Kadrey illustrate why a slogan is unreliable. ROSS involved a competing legal-research product, copied headnotes and a record on market substitution. Kadrey involved different works, a different system and a different summary-judgment record. The Copyright Office's Part 3 report also treats purpose, source, amount and market effects as context-sensitive questions rather than one universal outcome.
For a buyer, the defensible move is to preserve the actual use: which content enters the service, whether it is licensed or client-supplied, what retrieval or fine-tuning occurs, what output is delivered, and whether the workflow creates a substitutive product. Counsel can then assess the relevant law without starting from a vendor slogan.
Read the customer contract beside the cases
| Contract field | Evidence to preserve | Common gap |
|---|---|---|
| Input rights | Customer representation, permitted sources and uploader authority | The firm assumes provider terms license the firm's source material. |
| Output rights | Assignment language, territorial limits and similar-output clause | Ownership language is treated as a non-infringement warranty. |
| Indemnity | Covered claim, remedy, notice, control and exclusions | Protected output is assumed to include fine-tuning, modified output or prohibited use. |
| Third-party features | Connector, model, retrieval source and incorporated terms | The base provider's protection is copied to another service. |
| Beta and preview | Separate warranty and indemnity exclusions | Production protection is assumed for a preview route. |
| Change control | Version, effective date, notice and exit rights | A current webpage replaces the executed agreement without comparison. |
Anthropic's commercial terms, OpenAI's services agreement and Google's Gemini API terms use their own structures. Check the executed route and order form on September 3, 2026; do not splice the strongest sentence from three providers into one imaginary contract.
Control the inputs and outputs the firm chooses
- Classify approved source material: public, licensed, client-provided, firm-owned or restricted.
- Record who has authority to upload or retrieve it.
- Keep source attribution where the workflow needs it.
- Review outputs for close reproduction, embedded third-party material and intended commercial use.
- Escalate high-volume transformation, fine-tuning and productization separately from ordinary assisted drafting.
- Retest when the model, retrieval corpus, terms or output use changes.
Separate acquisition from training and output use
A useful copyright file does not reduce the system to one verb. Record how the provider or customer obtained the works, how copies were prepared, what training or retrieval process used them, what the resulting model retains, and how the buyer uses the output. A court can treat one stage differently from another. The same source can also be licensed for one purpose and restricted for another.
For customer-controlled retrieval or fine-tuning, identify the corpus owner, license scope, territory, users, reproduction rights, machine-processing permissions and termination effect. For provider training, ask what the public and executed terms say, which model or service is covered, and whether an opt-in, feedback channel or product tier changes the default. Preserve separate answers instead of calling the whole route “licensed” or “unlicensed.”
Describe the works and market with precision
Copyright analysis changes with the material. Editorial headnotes, books, software, photographs, client documents and public-domain sources do not create the same record. Note whether the work contains protected expression, factual material, code, confidential information or contractual restrictions. Record the quantity and selection method rather than relying on the size of the overall dataset.
Market evidence also needs a defined product and customer. Ask whether the model or output substitutes for access to the original, supports a different internal task, competes with a licensed derivative market, or merely helps locate the source. Do not infer market harm or absence of harm from model scale. Link the claim to the evidence the cited court actually considered.
Use a proportionate output review
Not every generated sentence needs a copyright investigation. Review becomes more important when the prompt requests continuation, imitation, extraction, transformation of a known work, or production of a market-facing asset. It also matters when the output contains unusually specific language, citations to source passages, recognizable characters, code or images that appear close to existing material.
Set a stop rule for potential close reproduction. Preserve the prompt, source context, output, model and reviewer where retention policy permits. Do not publish the questionable material while provenance remains Unknown. For ordinary legal drafting, keep approved precedent and licensed sources in firm-controlled systems and require the lawyer to review substantive language rather than treating generation as a clean-room process.
Test indemnity against the real workflow
Build a small scenario table from the executed clause. Test a claim involving ordinary prompted output, customer-supplied content, retrieved third-party material, fine-tuning, a connector, modified output and continued use after notice. Record whether the provider controls defense, which remedies are available, whether the customer must stop use, and which conduct triggers an exclusion.
The purpose is not to predict litigation. It is to reveal when the operational design steps outside the purchased protection. If the contract excludes output created to imitate a third party, the product should not invite that prompt. If customer-provided data sits outside the coverage, the source-approval control becomes more important. Align the interface, policy and review process with the clause.
Monitor documents, not the news cycle
Assign owners to the litigation record, vendor terms and approved workflow. For cases, monitor the docket for appeal, amended order, class decision or settlement change. For contracts, record the effective date, incorporated policy pages and notice mechanism. For the workflow, trigger review after a model, retrieval source, feature or output use changes.
Store the prior state and exact delta. A new complaint may raise a credible issue without changing a holding. A new order may decide one claim and leave others open. A terms update may improve one indemnity clause while narrowing an eligible product. The change note should say what moved, what did not, and which approved workflows are affected.
Give decision-makers a one-page copyright record
Summarize the approved service, input classes, source authority, output uses, contract version, relevant indemnity, prohibited conduct, named cases, procedural dates, reviewer and next check. Put Unknown beside evidence the buyer cannot obtain. Link the detailed research file instead of copying every allegation into the executive note.
The resulting decision can be narrow: approve public-source orientation with human review; restrict uploaded copyrighted corpora; negotiate protection for a production use; or hold the workflow pending evidence. This is more useful than declaring the provider safe or unsafe as a whole.
What a buyer can decide now
A buyer can choose a route with clearer terms, restrict source classes, preserve provenance, require output review and negotiate allocation for the actual workflow. The buyer cannot turn an unsettled legal landscape into certainty by adding a disclaimer or selecting a famous provider. The premium control is a short evidence file that makes the remaining uncertainty visible and assignable.
FAQ
Have US courts decided that AI training is fair use?
No universal rule follows from the current record. ROSS and Kadrey reached different results on different facts and records. Official source · checked 2026-09-03
Did the Bartz settlement decide whether Anthropic infringed copyright?
No. The July 20, 2026 order approved a class settlement; settlement approval is not a merits holding that all conduct was lawful or unlawful. Official source · checked 2026-09-03
Does an output-ownership clause eliminate copyright risk?
No. Ownership language, non-infringement warranties and indemnity are different contract fields and may carry exclusions. Official source · checked 2026-09-03
Sources checked
- Copyright and AI initiative, checked 2026-09-03.
- Part 3: Generative AI Training, checked 2026-09-03.
- Thomson Reuters v. ROSS opinion, checked 2026-09-03.
- Kadrey v. Meta order, June 25 2025, checked 2026-09-03.
- Bartz final settlement approval, July 20 2026, checked 2026-09-03.
- Commercial Terms, checked 2026-09-03.
- Services Agreement, checked 2026-09-03.
- Gemini API Terms, checked 2026-09-03.
Operational information, not legal advice. Verify current terms, account configuration and applicable professional duties before use.