These records mirror Damien Charlotin’s AI Hallucination Cases Database. This page separates what we import from what we add and from the other sources we touch, and every item names the file that does it. Corpus last imported Oct 20, 2026.
All 2,202 case records come from the database’s CSV export, downloaded and re-published automatically. The cases, their facts, outcomes and source links are Damien Charlotin’s research and the work of the contributors he credits. We did not find these cases; we mirror them and add the layers below. Fields taken from the export:
License: CC BY 4.0. Reuse requires credit to the creator, a link to the license and a note of changes; this page is that note. Evidence: scripts/import-charlotin-csv.mjs, data/meta.json.
Each item exists in the repository today. “Method” says whether it is rules, software or a language model, and whether a person reviewed it. None of this is Damien’s work or opinion, and none of it implies his endorsement.
Country codes (USA to US, UK to GB), US state codes, penalty amounts as numbers, URL-safe record ids, and the federal / state / international level and US circuit derived from the court name.
Keyword rules over the imported outcome and hallucination text assign failure tags (fake citations, fabricated quotes, misrepresented authority), sanction types, control gaps, and a four-level editorial impact (low to career-ending). The impact level is AI Vortex's own heuristic, not Charlotin's classification and not a legal finding.
Re-imports from the database, links a renamed or re-dated record to the same URL so ids and first-seen dates survive, reports what was added, removed or corrected upstream, and fails a freshness check when the corpus is more than 48 hours old.
The database records tools as free text (“Claude AI”, “MS Copilot”, “GPT-4o”). A canonical vocabulary of 21 tools groups those variants so filters, counts and tool pages do not split the same product. Stored values are untouched; the mapping is applied when reading. Tools outside the vocabulary keep their raw text.
Searches accept court shorthand (D.N.J., S.D.N.Y.) and jurisdiction aliases (US, USA, United States), show why each result matched, and disclose a broader fallback instead of silently substituting results.
Each case link is classed as official court or government, docket mirror or free legal archive, publisher archive, or secondary report, from the link's host. It says where the link points. It is not a verification of the case.
A language model (Gemini 2.5 Flash through the KIE API) reads the linked document and drafts a structured summary: judicial reasoning, decision context, practical implications and an evidence status. 1,621 of 2,202 records have extracted judicial reasoning; 1,599 are marked primary-document-verified by that automated pass. These are AI-drafted, not human-reviewed (human-reviewed records: 0). Some fields are template text built from the imported record.
The database does not record judges. Where the linked opinion names the deciding judge, a deterministic extractor stores the name with the page locator and a quote. This covers 54 of 2,202 records, so judge pages are partial.
902 generated profile pages group the records by those dimensions, with counts, related entities and print-ready reports. Pages below an evidence threshold are kept out of search indexes.
Counts by year, country, court and tool; an interactive map; and related-case links scored from shared court, state, tags and tool. Counts describe the tracked corpus, not usage-adjusted rates.
JSON and CSV exports of the normalized records with a versioned manifest, record count and checksum, carrying the Charlotin attribution and license notice.
Lets ChatGPT, Claude, Cursor and other MCP clients search the corpus, pull case detail, build jurisdiction and tool-risk briefs and pre-filing packets. Every response carries the Charlotin attribution.
Per-case briefs, a pre-filing gate, a filing-integrity scanner, a control-maturity profile and a policy studio, plus printable and downloadable reports. These are AI Vortex tools built on the records.
A digest builder for cases added in the last week and a signup form. The builder only writes files; sending is a separate step.
| Source | Used for | License |
|---|---|---|
AI Hallucination Cases Database (Damien Charlotin), CSV exportscripts/import-charlotin-csv.mjs (download.csv); data/meta.json | The case corpus: names, courts, dates, parties, recorded AI tool, hallucination items, outcomes, penalties and the source link of every record. | CC BY 4.0 |
Documents linked from each record (court sites, dockets, publisher archives, news)data/cases.json (source_url); scripts/enrich-case-intelligence-kie.mjs | Opened from each case's source link; read by the automated document review and by judge extraction. The linked document controls over any summary here. | Rights stay with each publisher; not covered by Charlotin's CC BY 4.0 |
Wikimedia Commons and public-domain portrait sourcesdata/entity-media.json; scripts/sync-entity-media-assets.mjs | Judge portraits and court images on entity pages, each with credit, source page and license. | CC0, public domain, CC BY-SA (per image) |
world-atlas (Natural Earth) and us-atlas (US Census Bureau) npm packagespackage.json; components/SanctionsMapV2.tsx | Map and state outline geometry. | Packages are ISC-licensed; underlying geometry from Natural Earth and the US Census Bureau |
AI vendor product pageslib/tool-catalog.ts | Links on AI tool profiles, from a hand-maintained catalog. | Links only |
Gemini 2.5 Flash via the KIE API (processing service, not a data source)scripts/enrich-case-intelligence-kie.mjs | Language-model extraction over the linked documents to draft the case summaries described on the 'What AI Vortex adds' page. | Service; output is AI Vortex's, unreviewed by a human |
The 2,202 records link to 59 different sites. Most common: damiencharlotin.com (2,040), reason.com (37), law.justia.com (13), naturalandartificiallaw.com (13), v.daum.net (6), caselaw.findlaw.com (5), govinfo.gov (5), storage.courtlistener.com (5), casemine.com (4), gao.gov (4), leagle.com (4), courtlistener.com (3). The case page shows each record’s own source and the contributor the database credits.
