Study published October 21, 2026 · data snapshot 2026-10-21 · decisions dated to 2026-10-15
Courts recorded 358 matters with AI-fabricated citations or quotes dated July to September 2026, against 420 in the second quarter (-14.8%). That is the count as of today, three weeks after the quarter closed, and late entries are normally added, so read it as a floor. Across all 2,203 matters in Damien Charlotin's public database, 57.4% involve a pro se litigant and 39.6% a lawyer; 25.0% of lawyer cases carry a monetary sanction with a recorded amount, against 7.2% of pro se cases.
We recomputed everything from the raw database export. The script, the snapshot, the derived CSV and the method are on this page, so any figure can be re-run. This is an observational count of recorded court decisions: it does not measure how often lawyers or litigants use AI, and it does not explain why any number moved.
Five findings
- Volume. 2,203 matters in total; the peak month is March 2026 with 182. Q3 2026 months: July 136, August 121, September 101.
- Pro se vs lawyers. Pro se litigants are the majority and have been in every quarter since 2025 Q3 (57.0% in Q3 2026, 58.8% in Q2). Lawyers are 40.2% in Q3.
- Sanction type. The mix differs sharply by filer: 17.9% of lawyer matters carry a professional-sanction flag, against 0.2% of pro se matters.
- Jurisdictions. 1,502 matters (68.2%) are in the United States; 43 jurisdictions appear in total.
- Q3 vs Q2 2026. Fewer matters recorded (358 vs 420), a higher US share (76.5% vs 71.7%) and a lower share with a monetary amount (10.9% vs 15.2%). Part of the Q3 gap is database lag (see limitations).
1. Volume by month
Monthly matters rose from 16 in January 2025 to a peak of 182 in March 2026, then fell back to 129 in April, 155 in May and 101 in September. 861 matters are dated 2025 and 1,264 are dated 2026 so far (October is partial and left out of the chart).
| Quarter | Matters | Pro se | Lawyers |
|---|---|---|---|
| 2025 Q1 | 57 | 54.4% | 36.8% |
| 2025 Q2 | 124 | 50.0% | 46.8% |
| 2025 Q3 | 270 | 52.2% | 45.2% |
| 2025 Q4 | 410 | 59.5% | 37.3% |
| 2026 Q1 | 455 | 59.1% | 39.1% |
| 2026 Q2 | 420 | 58.8% | 37.9% |
| 2026 Q3 | 358 | 57.0% | 40.2% |
2. Pro se litigants vs lawyers
Of 1,265 pro se matters and 872 lawyer matters (the rest are judges, experts, arbitrators, mixed or unspecified), the pro se share has moved within a narrow band: 50.0% in 2025 Q2, 59.5% in 2025 Q4 and 57.0% in 2026 Q3. The lawyer share was 46.8% in 2025 Q2 and 37.3% in 2025 Q4. The table in section 1 has every quarter. "Lawyers" includes government lawyers and prosecutors; matters with several filer types are counted separately as mixed or other.
The database does not say how many lawyers or litigants use AI, so these shares describe who appears in recorded decisions, not who is more likely to hallucinate.
3. Sanction types
Each matter is assigned its most severe recorded consequence. A warning is the most common outcome (30.2% of dated matters) and 20.9% have no outcome recorded in the database. Where a US-dollar amount is recorded (193 matters) the total is $1,423,406, the median $2,000 and the largest $110,204; amounts in other currencies are not converted and not added.
| Filer | Matters with a date | Monetary amount | Professional flag | Warning most severe |
|---|---|---|---|---|
| Pro se litigants | 1,265 | 7.2% | 0.2% | 40.6% |
| Lawyers (incl. government lawyers) | 871 | 25.0% | 17.9% | 17.5% |
| Consequence | Matters | Share |
|---|---|---|
| Warning or admonishment | 665 | 30.2% |
| Outcome not recorded | 461 | 20.9% |
| Monetary sanction with an amount | 317 | 14.4% |
| Filing struck, ignored or case dismissed | 198 | 9.0% |
| Professional sanction (bar referral, reprimand, disqualification) | 146 | 6.6% |
| Show-cause or explanation order, pending | 143 | 6.5% |
| Costs or fine mentioned, no amount recorded | 93 | 4.2% |
| Other outcome | 91 | 4.1% |
| Remedial order (CLE, disclosure, certification) | 47 | 2.1% |
| Court recorded no sanction | 41 | 1.9% |
4. Jurisdictions
The United States accounts for 1,502 of 2,203 matters (68.2%); Canada (234), Australia (112) and the UK (76) follow. 43 jurisdictions appear in total, 18 of them in Q3 2026 and 21 in Q2. The ranking reflects where courts publish decisions and where the database's contributors look, so it is not a ranking of where AI misuse is highest.
| Jurisdiction | Matters | Share | Q2 2026 | Q3 2026 |
|---|---|---|---|---|
| USA | 1,502 | 68.2% | 301 | 274 |
| Canada | 234 | 10.6% | 52 | 29 |
| Australia | 112 | 5.1% | 17 | 20 |
| UK | 76 | 3.4% | 8 | 8 |
| Israel | 58 | 2.6% | 4 | 3 |
| Brazil | 41 | 1.9% | 9 | 2 |
| India | 16 | 0.7% | 3 | 4 |
| Italy | 15 | 0.7% | 4 | 0 |
| France | 14 | 0.6% | 2 | 3 |
| Germany | 11 | 0.5% | 0 | 0 |
5. Q3 vs Q2 2026
Q3 2026 has 358 matters against 420 in Q2. The US share is 76.5% against 71.7%: Canada went from 52 to 29 matters while Australia went from 17 to 20. The share of matters with a recorded monetary amount is 10.9% against 15.2%, and the professional-sanction flag is at 6.7% against 6.0%.
| Measure | Q2 2026 | Q3 2026 | Change |
|---|---|---|---|
| Matters | 420 | 358 | -14.8% |
| Pro se | 247 | 204 | -17.4% |
| Lawyers | 159 | 144 | -9.4% |
| Monetary sanction with an amount | 64 | 39 | -39.1% |
| Professional sanction flag | 25 | 24 | -4.0% |
| Jurisdictions with at least one matter | 21 | 18 | -14.3% |
| US matters | 301 | 274 | -9.0% |
We do not claim a cause for any change. Recording lag alone can move these numbers: the database added 455 Q1 2026 matters by today against 422 on April 24, 24 days after that quarter closed (7.8% more). Q3 2026 was measured 21 days after it closed, so Q3 will probably rise as entries arrive. If it grew by the same 7.8%, it would still sit below Q2. That is an inference from one quarter, not a forecast.
Method
- Source. AI Hallucination Cases Database by Damien Charlotin (CC BY 4.0), CSV export downloaded on 2026-10-21: 2,203 rows, 2,202 with a decision date, latest dated 2026-10-15. SHA-256 of the snapshot:
108ff0249c2a8c29...(full hash inhallazgos.json). - Periods. Quarters follow the decision date in the database (Q3 = July to September). Rows without a date are left out of time series and counted in totals.
- Filer. From the "Party(ies)" field: pro se = Pro Se Litigant; lawyer = Lawyer, Government Lawyer or Prosecutor; other = judge, expert, arbitrator and combinations; mixed = pro se with a lawyer.
- Most severe consequence. Order of precedence: monetary amount above 1 in the penalty field, professional sanction, costs or fine without amount, filing struck or case dismissed, remedial order, warning, pending order, none recorded, not recorded. Free-text outcomes are matched with keyword rules written in
analisis.py. - Jurisdiction. First country listed in the "State(s)" field; arbitration-only entries are grouped separately and not counted as a country.
- Reproduce.
python3 analisis.py(Python 3, standard library only) in the data folder writes every CSV andhallazgos.json. A build test checks that each figure on this page equals the CSV or JSON value.
Limits
- The database is a curated collection, not a census. It depends on what courts publish and what contributors find.
- Recent quarters are incomplete and are revised upward later.
- No denominator: the number of filings, lawyers or AI users is unknown, so none of these are rates of AI misuse.
- Sanction classes come from free-text outcomes and keyword rules; 20.9% of matters have no outcome and 4.1% fall in an unclassified "other" bucket.
- This page counts matters with a recorded US-dollar amount only. Our statistics page uses the AI Vortex normalized corpus (2,202 records, 43 jurisdictions, consequence tags), so its monetary count and country count are defined differently from this study.
- Correlation only: the filer split and the sanction split do not show that one group is treated more harshly for the same conduct, since the conduct and the court differ.
Download the data and the script
- casos_derivados.csv: one row per matter with normalized date, quarter, country, filer class and consequence class (plus the Charlotin source link)
- volumen_mensual.csv, trimestres.csv, sanciones.csv, sanciones_por_parte.csv, jurisdicciones.csv, tendencia_3t_vs_2t.csv
- hallazgos.json: every headline figure with the snapshot hash
- analisis.py and the source snapshot (CC BY 4.0, Damien Charlotin)
Cite and credit
Cite as: AI Vortex, "AI Hallucinations in Court, Q3 2026", https://www.aivortex.io/legal/ai-case-law/ai-hallucinations-q3-2026/, October 21, 2026. Data: Damien Charlotin, AI Hallucination Cases Database.
Credit: the underlying records come from Damien Charlotin's AI Hallucination Cases Database, licensed CC BY 4.0. AI Vortex normalized the filer, consequence and jurisdiction fields and computed every statistic here; those transformations are ours, not Charlotin's, and he has not endorsed this study. Check the primary court record before relying on any case. See also what AI Vortex adds in the case database notes.
The bottom line: recorded AI-hallucination matters are mostly pro se, mostly in US courts and mostly end in a warning; lawyers are the group that gets the monetary and professional consequences. Q3 2026 is recorded below Q2 so far, with a known lag that makes it a floor.
Related
AI-Assisted Research. The analysis was run by script on the 2026-10-21 snapshot and the page was drafted with AI assistance. Every figure is reproducible from the files above; corrections to manuel@aivortex.io.