Updated October 21, 2026 · computed from the public dataset, snapshot 2026-10-20
As of the 2026-10-20 snapshot, the AI Vortex corpus holds 2,202 court matters from 2023-04-14 to 2026-10-15 in which AI-generated or AI-implied false content appears in the record, across 49 countries. 1,499 are US matters (68%). The count rose from 61 matters dated 2024 to 861 in 2025, and 1,264 are already dated 2026 through 2026-10-15. 1,268 (58%) are tagged pro se, so the corpus cannot tell you how many lawyers were sanctioned: it has no attorney flag. It is a research layer built on Damien Charlotin's database, not a complete census.
The key numbers
| Measure | Value | What it means |
|---|---|---|
| Matters in the corpus | 2,202 | Court matters where AI-generated or AI-implied false content appears in the record, 2023-04-14 to 2026-10-15 |
| Countries | 49 | 1,499 matters (68%) are US; Canada, Australia and the UK follow |
| Tagged pro se | 1,268 (58%) | The other 934 are lawyer-filed or filer unclear; the corpus has no separate attorney flag |
| Fabricated citations | 1,845 | Matters tagged fake-citations; a matter can carry several tags |
| Misrepresented authority | 945 | Real authority cited for something it does not say |
| Fabricated quotes | 730 | Quotation not found in the cited source |
| Matters in the latest 30 days | 88 | Decisions dated up to 2026-10-15 |
Source: AI Vortex dataset JSON, version 2026-10-20. Tags overlap, so the failure-mode counts do not add up to the total.
How many AI hallucination cases are there per year?
861 matters are dated 2025 and 1,264 are dated 2026 so far. Earlier years are small: 16 in 2023 (the first dated 2023-04-14) and 61 in 2024. Part of the growth is real filing behaviour and part is better discovery by the people who build the underlying database, so read the slope as direction, not as a rate.
| Year | All matters | US matters | Coverage |
|---|---|---|---|
| 2023 | 16 | 11 | full year |
| 2024 | 61 | 38 | full year |
| 2025 | 861 | 527 | full year |
| 2026 | 1,264 | 923 | to 2026-10-15 |
Which countries and US states have the most cases?
The United States dominates the record with 1,499 matters. Outside it, Canada, Australia and the United Kingdom lead. This reflects where courts publish and where the underlying database looks, not necessarily where AI misuse is highest.
| Country | Matters | Share of all |
|---|---|---|
| United States | 1,499 | 68% |
| Canada | 233 | 11% |
| Australia | 112 | 5% |
| United Kingdom | 76 | 3% |
| Israel | 58 | 3% |
| Brazil | 41 | 2% |
| India | 16 | 1% |
| Italy | 15 | 1% |
| France | 14 | 1% |
| Germany | 11 | 0% |
| State | Matters | Share of US |
|---|---|---|
| California | 154 | 10% |
| New York | 134 | 9% |
| Texas | 89 | 6% |
| Florida | 74 | 5% |
| Illinois | 71 | 5% |
| Arizona | 55 | 4% |
| Michigan | 54 | 4% |
| Pennsylvania | 45 | 3% |
| Washington | 42 | 3% |
| Indiana | 38 | 3% |
What happens to lawyers and litigants who file AI-fabricated cases?
Most matters end in a warning or no recorded sanction. 459 carry a monetary sanction and 189 a bar referral. Among US matters with a recorded dollar amount (228), the amounts total $1,524,485, the median is about $1,606 and the largest is $110,204. Some entries record nominal amounts, so treat the median as indicative.
| Consequence | Matters | Meaning |
|---|---|---|
| Warning | 711 | Court cautioned the filer without further penalty |
| Monetary sanction | 459 | 228 US matters carry a recorded dollar amount |
| Bar or disciplinary referral | 189 | Referral to a bar authority |
| Professional consequence | 168 | Disciplinary or professional-standing outcome as normalized in the corpus |
| Filing struck | 99 | Brief or motion struck |
| Case dismissed | 122 | Matter dismissed |
| Order to show cause | 106 | Court asked the filer to explain; outcome may be pending |
| No adjudicated consequence recorded | 728 | No sanction recorded in the source |
Which AI tools show up in the cases?
Mostly none by name: 1,959 matters (89%) do not identify the tool. Where a tool is named, the count is: ChatGPT 115; Copilot 5; CoCounsel 4; Claude 4; Google AI 3; Perplexity 3. A named tool is only what the court record or filer disclosed, so this is not a vendor failure rate. For rates on legal research products, see the tracker comparison and the legal research tools page.
Method, limits and how to cite
The numbers are computed by scripts/aeo-data-stats.py from the public dataset (version 2026-10-20, 2,202 records, upstream snapshot hash b90cfbd48004). The records mirror Damien Charlotin's AI Hallucination Cases Database (CC BY 4.0), which AI Vortex imports, normalizes and classifies; AI Vortex is not endorsed by or partnered with Charlotin. Counts describe this corpus, not usage-adjusted incidence. Case summaries do not replace the linked order or docket.
Cite as: AI Vortex, Legal AI Risk Corpus, snapshot 2026-10-20, https://www.aivortex.io/legal-ai-risk/dataset. Search every matter in the case database, see the live analytics and the methodology. Report an error through the corrections form.
How to choose
- Need the live number? Use the case database; this page is a dated snapshot.
- Need to know how a court treated a specific filing? Open that matter and read the linked order.
- Need the tool-level story? Treat named-tool counts as disclosure, not failure rates.
- Need the rules? See judge orders and AI disclosure requirements in the guides linked below.
The bottom line: 2,202 matters in 49 countries, mostly US, mostly pro se, mostly warnings. The data supports 'this keeps happening and keeps growing in the record', not a per-lawyer incidence rate.
Sources and what we checked
Everything here was computed on 2026-10-21 from the public dataset and can be re-run.
- AI Vortex Legal AI Risk dataset (JSON), version 2026-10-20, 2,202 records
- Dataset page: version, license and citation guide
- AI Hallucination Cases Database, Damien Charlotin (CC BY 4.0, upstream source)
- AI Vortex sanctions tracker article (earlier snapshot; figures differ by date)
Related decision paths
AI-Assisted Research. This page was researched and written with AI assistance. Prices and product facts were computed from the public dataset (snapshot 2026-10-20) on 2026-10-21; where a figure could not be verified it is left out. Vendors can request corrections at manuel@aivortex.io.