A public database of court decisions in which someone filed AI-invented case citations
1,983 decisions as of 30 August 2026. You cannot build the database, but you can adopt the one-line verification rule that every entry in it documents the absence of.
AI Hallucination Cases Database
Global
2023–20264 minutes
- Legal work and accountability
- Researcher or analyst
- Research
- Drafting reports and advocacy material
What exists
Damien Charlotin, a research fellow at HEC Paris, maintains a free, downloadable, public-domain database of legal decisions in which a court found, not merely alleged, that someone relied on AI-hallucinated content, usually fabricated case citations. Each entry records the jurisdiction, the court’s finding, the sanction, and the tool where one was identified. His definition is why the dataset is worth anything:
Notably, this does not cover mere allegations of hallucinations, but only cases where the court or tribunal has explicitly found (or implied) that a party relied on hallucinated content or material.
As of 30 August 2026 it lists 1,983 decisions: 6 in the second quarter of 2023, 55 in the first quarter of 2025, 403 in the fourth quarter of 2025, 446 in the first quarter of 2026, then a decline. Courts have struck briefs, imposed five-figure fines and referred lawyers to disciplinary boards. In one United States appellate case two attorneys were sanctioned fifteen thousand dollars each.
1,983 is a floor on detected cases, not a rate. Three entries come from the entire Arab world and none from Iraq, which tells you about court publishing and English indexing, not about Iraqi filings. Do not read that silence as safety.
Nor is it a lawyers’ problem. A 57% majority of entries are self-represented people who turned to a chatbot because they could not afford counsel: the resourcing profile of a small organisation with no research staff.
The rule to adopt on Monday
No engineer, no budget, no software.
- Adopt the court’s own rule as an office rule. No document leaves the office containing any citation, whatever its source, that a named person has not personally opened and read. Being source-agnostic makes it easy to explain to non-technical colleagues and impossible to argue with.
- Make verification a named step with a named owner. Before any footnoted document ships, a shadow report to a treaty body, a grant application, a research brief, a legal submission, one person opens every cited source. Give it to someone who did not draft the document.
- Verify by opening the source, never by asking the assistant to confirm it. “Are you sure?” generates a new answer; it does not check the old one. The database’s own chart of excuses is full of people who did exactly that.
- Keep a one-line verification log per document: who checked, which sources, what date. That turns “we are careful” into evidence when a funder, a court or a hostile government challenges the document a year later. For an Iraqi organisation submitting to international bodies, that audit trail is the actual protection.
- Use the dataset as free training material. One click, no account, into any spreadsheet. Filter it to a jurisdiction or a field of law and read the real excuses and the real sanctions: a staff training session built out of primary court records, at no cost.
What this does not tell you
Nothing here says which assistant is likelier to invent a citation: 88% of entries name no tool. The lesson is not about a product. Any document with footnotes can carry invented sources that look exactly like real ones, and the only thing standing between your organisation and that outcome is a person who opened the link.
Tools the sources name
The sources name no tool for this case. We have not guessed one, and neither should anyone repeating the story.
What we do not claim
Everything above is limited by what its sources actually prove. This is the part they do not.
Every figure on this page is as of 30 August 2026. The database updates continuously, so date-stamp any number you quote from it.
Do not say the trend is still rising. It plateaued and turned down after the first quarter of 2026, and the maintainer himself uses the word "plateau". Claiming an accelerating crisis when the person holding the data says otherwise is the kind of overreach a hostile reader finds in ten minutes.
Do not frame this as a lawyers' problem. A 57% majority of the entries are people representing themselves, often because they could not afford a lawyer.
Do not attribute the hallucinations to ChatGPT or any other named product. 88% of entries name no tool at all, and the database explicitly disclaims tool causation. Naming a vendor here would be both unsupported and legally reckless.
Do not present 1,983 as an incidence rate, or as a measure of how often AI invents things, or as a trend in model quality. It counts decisions where a court both noticed and said so in writing, overwhelmingly in English-language common-law systems: the United States alone is 69%. The maintainer's own FAQ concedes it is necessarily an undercount.
Do not imply we know anything about Arabic-language courts. There are three entries across the entire Arab world, and none from Iraq. That is an absence of court-published, scraped, English-indexed data. It is not evidence that the problem is absent here.
Do not present the maintainer's own citation-checking tool as something readers can obtain or use. It is demo-on-request, with no pricing and no independent evaluation, and we do not vouch for it.
Building the database is not the replicable action. The verification discipline is.