Not all of them are sanctions. Many are warnings. But 131 carry a professional sanction, 319 carry a monetary penalty, and 722 involve a practising lawyer rather than a self-represented litigant.
Each bar is the number of decisions dated to that quarter. The rise reflects both wider AI use and courts actively looking for fabricated authority in a way they were not two years ago.
Showing the most recent quarters. The final bar is partial, covering only the portion of the quarter captured at export. Because the database records only decisions where a court explicitly found or clearly implied reliance on hallucinated material, it undercounts: filings that were never caught do not appear.
A single decision often records more than one kind of defect, so these categories overlap. Fabricated authority is the failure mode that gets reported. Misrepresentation, where a real case is cited for something it does not hold, appears in 772 decisions and is far harder to catch by eye.
| USA | 1,278 |
| Canada | 204 |
| Australia | 97 |
| UK | 61 |
| Israel | 55 |
| Italy | 15 |
| contract | 433 |
| administrative | 248 |
| civil rights | 190 |
| employment | 168 |
| tort | 160 |
| other | 129 |
Median penalty where one was imposed is $1,500. The distribution has a long tail, and the tail is where fee-shifting and opposing-counsel cost awards live rather than flat Rule 11 fines.
| Case | Court | Date | Party | Penalty |
|---|---|---|---|---|
| Arabyads Holding Limited v. Gulrez Alam Marghoob Alam | ADGM | Dec 2025 | Lawyer | $282,508 |
| Couvrette v. Wisnovsky | D. Oregon | Mar 2026 | Lawyer | $110,204 |
| ByoPlanet International v. Johansson and Gilstrap | S.D. Florida | Aug 2025 | Lawyer | $85,567 |
| LiveVideo.AI Corp. v. Redstone | S.D. New York | Jun 2026 | Lawyer | $80,056 |
| Minjie Zheng v. ICANN | C.D. California | Dec 2025 | Pro Se Litigant | $66,129 |
| Howell Management Services, LLC v. Vashisht-Rota | CA California | Dec 2025 | Pro Se Litigant | $64,235 |
US-dollar penalties only. Awards in other currencies are recorded in the source database and are excluded here rather than converted.
The database records a named AI tool in only 210 decisions. In the rest the court recorded the use as implied or unidentified. Courts are not, in general, able to establish which system produced the fabrication, and neither is the filer after the fact.
When a filing is challenged, the question is not which model was used. It is what the filer did to check the output before it entered the record. Without a contemporaneous artifact, that answer is a recollection.
VariLegal produces a verification certificate at the time of checking, backed by evidence, provenance, and a SHA-256 integrity hash, independent of the AI system that generated the original output.
Every figure on this page is computed from the AI Hallucination Cases Database, compiled and maintained by Damien Charlotin, research fellow at HEC Paris. The database is available at damiencharlotin.com/hallucinations and is updated continuously.
Its inclusion rule is strict: it records decisions where a court or tribunal explicitly found, or clearly implied, that a party relied on hallucinated material. Mere allegations are generally excluded. It does not attempt to capture the wider universe of fabricated citations that were never addressed by a court.
This page reflects an export covering April 2023 through 4 August 2026, comprising 1,847 decisions. Counts shift as new decisions are added, so the source database is authoritative over anything shown here. Category counts overlap where a decision records more than one kind of defect. VariLegal is not affiliated with the database or its maintainer.
VariLegal verifies AI-assisted legal work product against authoritative sources and produces an independent certificate before the document enters the record.
How VariLegal verifies