Research: AI Hallucinations in Court — Damien Charlotin's Database and What It Counts
More than 2,100 court decisions worldwide finding that a filing relied on AI-invented citations, catalogued by a Paris lawyer and lecturer. Who runs the database, how it decides what counts, what the October 2026 figures show (pro se litigants outnumber lawyers; most cases are American; 2026 already has more than all prior years), and the anchor cases from Mata v. Avianca to a $110,204 sanction in Oregon.
Contents
Companion pages: Chatbot Liability and AI Agent Incidents.
The scale of AI-invented citations entering courts is usually stated with one source behind it: the AI Hallucination Cases database. This page sets out who runs it, how it is built, what it measures, and what it shows as of October 2026, so that any number taken from it carries its date and its limits.
1. Who runs it
Damien R. Charlotin is a French lawyer, legal researcher and lecturer based in Paris. His own CV lists him as a senior research fellow at HEC Paris (since January 2023), a lecturer at HEC Paris and Sciences Po teaching legal data analysis and a course on “Large Language Models and the Future of the Legal Profession,” an independent practitioner in public international law and arbitration (Pelekan Data Consulting), and a long-time contributor to Investment Arbitration Reporter (Charlotin, CV; homepage). The Financial Times described him in June 2026 as “a 34-year-old French lawyer and academic” (his media log). He started the database in April 2025 while teaching a Sciences Po course on the limits of language models in law (Artificial Authority FAQ), and maintains sibling databases on AI evidence and AI discovery (AI Evidence).
2. What counts
In his words, the database “tracks legal decisions … in cases where generative AI produced hallucinated content – typically fake citations, but also other types of AI-generated arguments. It does not track the (necessarily wider) universe of all fake citations or use of AI in court filings.” It “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,” with some exceptions he calls “a judgment call on my part” (the database).
Cases come from “a mix of referrals from benevolent randos on the internet … use of dedicated scrapers and bots to automatically monitor some data sources … and good old database searches with, wait for it, keywords.” He leaves the call to the courts: “I am not making that judgment, I let the courts and judges make or imply it, which is why the database is necessarily an undercount.” And: “most hallucinations are very obvious, making up a case name or a false quote is not, cannot be a human mistake” (FAQ). He notes the US skew: American court data is far easier to search than European.
3. What it shows, October 2026
The page reads “2184 cases identified so far” (last updated October 10, 2026). From the database’s own CSV (download), tallied October 14:
| Measure | Figure |
|---|---|
| Decisions by year | 2023: 16 · 2024: 61 · 2025: 859 · 2026 so far: 1,248 |
| By country | USA 1,492 · Canada 232 · Australia 112 · UK 76 · Israel 57 · Brazil 41 · others |
| Who filed | entries involving self-represented litigants 1,259; involving lawyers 868 |
| Tool named | ChatGPT in 114; the tool is “implied” or unidentified in most entries |
| Monetary penalty | 381 entries |
| Professional sanction | 164 entries |
Three things follow. The count is accelerating: 2026 has already produced more decisions than every earlier year together. The phenomenon is not mainly a lawyers’ scandal: self-represented litigants appear in more entries than lawyers do. And sanctions are the exception: most decisions end in a warning or a struck filing, not a fine.
Every total is a floor with a date. The July 8, 2026 snapshot was 1,731; in June the FT reported it had passed 1,500 from under 700 in the spring.
4. The anchor cases
- Mata v. Avianca (S.D.N.Y.), the first well-known case: attorneys Steven Schwartz and Peter LoDuca cited six decisions ChatGPT invented; Judge P. Kevin Castel imposed $5,000 in sanctions on June 22, 2023 (CourtListener docket).
- United States v. Cohen (S.D.N.Y., March 20, 2024): Michael Cohen got three non-existent cases from Google Bard, which he “did not realize … was a generative text service,” and passed them to his lawyer. Judge Jesse Furman called the citation “embarrassing and certainly negligent, perhaps even grossly negligent. But the Court cannot find that it was done in bad faith,” and found “no basis to question Cohen’s representation that he believed the cases to be real.” No sanctions (opinion).
- Couvrette v. Wisnovsky (D. Or., March 23, 2026): $110,204 against a lawyer, the largest US monetary penalty in the database and the “Oregon winery case” Charlotin told the FT was the highest fine he had seen.
- Seventeen in one day: Eugene Volokh, using the database, counted 17 US decisions noting suspected AI hallucinations issued on March 31, 2026 alone (Reason / Volokh Conspiracy, Apr 6 2026).
5. How far to trust it
It is the best-organized index of the phenomenon and is cited by courts and the press (Forbes, Jul 18 2025). Its strengths: the unit is a court finding, not an allegation; the judgment is left to judges; the data is open and filterable. Its limits: it is an undercount by design, US-skewed, one maintainer’s work, and the tool is not recorded in most cases, so “ChatGPT caused X” claims hold only for the named subset. Cite it as a dated lower bound.