Skip to content

reportAI HallucinationsAI in CourtsAI Governance

Legal AI incident report, Q3 2026 to 3 September: 54 incidents, 33 in US courts, 20 regulator outcomes

1 July to 3 September 2026: 54 tracker incidents in 11 jurisdictions, USD 132,888 in stated US penalties, 20 rows with a regulator referral or disposition.

Edited and verified by Cognesio LLP

Researched with AI assistance · sources verified by Cognesio LLP · How this was made ↓

As of 3 September 2026 the SafeLegalAI Global Legal AI Incident Tracker holds 54 incidents dated in the third quarter of 2026 (1 July to 3 September; the quarter is not over, and this report covers it only to today), against 8 in the second quarter and 11 in the third quarter of 2025. Thirty-three of the 54 are in United States courts, federal and state; the other 21 are spread across Canada, India, Brazil, Australia, Israel, South Africa, the United Kingdom, the Netherlands and China.

The tracker stood at 122 rows this morning after a sweep added 50 rows and updated 8, so the quarter’s count reflects new decisions and backfill together, not a rise in the rate of misconduct. Of the 54 rows, 47 are verified against a primary document and 7 are flagged unverified. Every number below is computed from the incident records by a script saved in the run log, and the same data is exported at /tracker/incidents.json.

How many incidents were recorded in Q3 2026?

PeriodAll rowsVerifiedUnverified
Q3 2026 (1 Jul to 3 Sep)54477
Q2 2026 (1 Apr to 30 Jun)871
Q3 2025 (1 Jul to 30 Sep)11110
Tracker total, 22 Jun 2023 to 2 Sep 202612210913

By month: 27 rows dated July, 26 dated August, 1 dated September so far. The quarter-on-quarter change is +46 rows and the year-on-year change is +43 rows, and neither figure should be read as a growth rate. The second quarter was never swept at the intensity applied to July and August; the 50 rows added today came from a single pass over decisions dated 10 July to 3 September. A fairer baseline is Damien Charlotin’s AI Hallucination Cases database, which held 2,009 entries on 3 September 2026 with 200 dated between 10 July and 3 September. Against that count the tracker’s 54 is a selection, not a census.

What the tracker adds is the layer after the courtroom: 42 of its 122 rows carry a recorded regulatory or disciplinary outcome, and 20 of those are third-quarter rows.

Where did they happen?

JurisdictionQ3 2026Q2 2026Q3 2025All rows
US federal202130
US state133432
Canada40110
India4018
Brazil3004
Australia2015
Israel2003
South Africa2005
United Kingdom23317
EU (Netherlands)1001
Other (China)1001

Eleven jurisdictions appear in the quarter; fifteen appear in the tracker as a whole. The US share is 61 percent of the quarter (33 of 54) and 51 percent of all rows (62 of 122). Charlotin’s database shows the same skew at larger scale: 163 of his 200 window entries are American. The UK, the tracker’s deepest jurisdiction with 17 rows overall, contributed 2 this quarter: the Crown Prosecution Service’s two non-existent extradition authorities in Tobosaru v Court of Law Craiova, and a litigant in person’s doctored documents in Lodhia v Twelve Trees.

Two entries are firsts for the tracker rather than for the world. The Dutch bar disciplinary council decision in ‘s-Hertogenbosch is the first EU row, and the Yichang district court admonishment is the first Chinese row, recorded from a court WeChat report and flagged unverified because the decision itself has not been located.

Who misused the AI, and with what?

ActorQ3 2026 rows
Lawyer36
Litigant in person9
Other (adjudicator or official)4
Judge1
Not coded4

The actor field was added to the schema with today’s sweep, so only third-quarter rows carry it (50 of the 54) and no cross-quarter comparison is possible yet. Within the quarter, lawyers account for two thirds of the rows. The five rows coded judge or other are the pattern the actor field first makes countable (two earlier Indian rows, Buckeye Trust in December 2024 and KMG Wires in October 2025, recorded an adjudicator’s own order resting on non-existent authority but carry no actor code): an acting judge in Johannesburg, a municipal disciplinary-hearing chairperson whose ruling reached the Labour Court in Durban, an income-tax appellate officer in Jaipur, an additional commissioner of customs in Surat and a state tax officer in India, each found to have rested a decision on authorities that do not exist.

AI tool named in the recordQ3 2026All rows
Not named in the record4172
ChatGPT723
Other single product (GPT-4o, LEAP, Visto.ai, Legal Genius, Microsoft Copilot, Google AI, MX2.law)07
More than one product named23
Generative AI, product not identified112
Claude12
Westlaw AI tools12
Strongsuit11

In 41 of 54 rows (76 percent) the judgment or order does not identify the product. The tracker never infers one. Where a tool is named it is usually because counsel admitted it: ChatGPT in seven rows, Claude with other platforms in Kleyman, Westlaw’s AI tools in Rosslyn2016, where counsel specifically denied using ChatGPT. Across all 122 rows, ChatGPT is named in 23 and every other product in two or fewer.

What did courts do?

OutcomeQ3 2026Q2 2026All rows
Sanctions15129
Other (order set aside, admonishment, self-report)11123
Fine709
Warning6214
Costs order5011
Referral5118
Suspension213
Pending203
Dismissal1212

Monetary penalties are summed only over rows where the order states a figure. Twenty of the 54 rows do.

CurrencyRows with a figureSumLargest
USD14132,88846,511 (Kleyman Law Group v Kaloidis)
CAD331,00020,000 (Arbuckle v Tanner)
AUD19,097.939,097.93 (Asif v Minister for Immigration)
ILS16,0006,000 (Haifa Labor Court)
EUR11,2501,250 (Raad van Discipline ‘s-Hertogenbosch)

The USD figures mix sanctions payable to a court fund, fee-shifting to the opposing party and reimbursements. The largest, USD 46,511 in Kleyman, is a USD 10,000 sanction plus USD 36,511.35 in fees and costs. The USD 14,500 in Full Standard Properties v SJ 6070 is a reimbursement counsel paid before the show-cause hearing, after which the order was discharged. The median stated USD penalty in the quarter is USD 3,250.

Five rows carry the quarter’s weight, and each is traced to its primary document:

  1. Kleyman Law Group v Kaloidis (Supreme Court of New York, Kings County, 4 August). A lawyer suing a former client’s estate filed 19 submissions citing 23 fictitious decisions and misstating at least 83 real ones, drafted with Claude and other tools. Sanction and fees of USD 46,511; referral to the Grievance Committee.
  2. Scott v Illinois Human Rights Commission (Appellate Court of Illinois, 28 July). Ten false items in two briefs, attributed by counsel to a ChatGPT subscription. A USD 15,000 fine set at USD 1,500 per misstatement, with the court stating that fines must rise until they deter; the clerk was directed to send the opinion to the Attorney Registration and Disciplinary Commission.
  3. FJL v TGO (High Court of South Africa, Gauteng Division, full court, 31 July). At least eleven fictitious or garbled citations in an acting judge’s first-instance judgment. Opperman J forwarded the judgment to the Legal Practice Council, reasoning that acting judges remain practising advocates within its jurisdiction.
  4. Gadiya v Union of India (Supreme Court of India, 2 September). A customs penalty order of INR 4.25 billion, and the High Court judgment upholding it, set aside because the order relied on AI-hallucinated precedents. The court extended its July ruling in Pooja Ramesh Singh from tribunals to executive adjudicators and left discipline to the officer’s appointing authority.
  5. In re Brian E. Mitchell (USPTO Office of Enrollment and Discipline, final order 27 July, published 27 August). A public reprimand by settlement for a patent attorney whose AI-drafted claim-construction chart cited non-existent parts of the patent’s own record. It is the only US third-quarter row in which a regulator, rather than a court, produced the final document.

What did regulators do?

Twenty third-quarter rows carry a regulatory outcome. Sixteen are referrals or notifications with no disposition yet on record; four are dispositions, and two of those four are unverified.

RowBodyWhat the record says
In re Brian E. MitchellUSPTO Office of Enrollment and DisciplinePublic reprimand by settlement, executed 20 and 21 July, final order 27 July, published 27 August
Raad van Discipline ‘s-Hertogenbosch 2026:93Dutch bar disciplinary councilFormal reprimand for breaches of expertise and integrity; EUR 1,250 costs
Lee v Law Society of OntarioLaw Society Tribunal (Ontario)Six-month suspension and CAD 10,000 costs; unverified, CanLII not retrievable
In re Faisal AhmedMinnesota Supreme Court / OLPR30-day suspension by stipulation, reported 21 August; order not located, unverified
Kleyman Law Group v KaloidisAttorney Grievance Committee (New York)Referred by the court; no disposition
Scott v Illinois HRCIllinois ARDCOpinion sent to the ARDC; no disposition
FJL v TGOLegal Practice Council (South Africa)Acting judge referred; no disposition
Lodhia v Twelve TreesBar Standards BoardDirect-access counsel referred; no disposition
Akerlund v Atlas Air, Parnell v Florida DOCEleventh Circuit Committee on Lawyer Qualifications and ConductCounsel referred in each; no disposition
JMOR Properties v Artist Alley, Capital Standard v US BankThe Florida BarCounsel referred in each; no disposition
Forney v Township of CaryNorth Carolina State BarCounsel referred; no disposition
NM Supreme Court (Aarons)New Mexico Disciplinary BoardReferred per press report; unverified
Barteca v Tacobarn, Tov Realty v SuarezConnecticut disciplinary authority / Statewide Grievance CommitteeReferral in one; courtesy copy of the order in the other
Cartagena v Dixon BlackburnSDNY Grievance CommitteeAdded to an existing referral; unverified
TJPR 0035599-44, TJSP 1032884-04, TRT20 SergipeOrdem dos Advogados do Brasil (state sections and Federal Council)Court notices sent to the bar; no disposition; the TRT20 row is unverified

The gap between referral and disposition is the tracker’s standing finding. In the whole record, 42 rows carry a regulatory outcome and most are still an open referral. The two verified dispositions this quarter came from a Dutch council and a US patent office, not from the bodies that received the most referrals. As of 3 September 2026 the tracker records no verified disposition from the Solicitors Regulation Authority, the Bar Standards Board or any US state bar on a third-quarter row; the Minnesota suspension is the nearest, and its order has not been located.

What changed in the rules?

The Regulation Tracker holds 84 documents as of 3 September 2026. Seven have a current version dated in the quarter: four new documents and three revisions.

DocumentJurisdictionNew or revisedRequirements coded
Misuse of AI: warning notice (SRA, 17 Aug)UKNewVerification, confidentiality, supervision, competence
Practice Direction HC 142 (High Court of Ireland, signed 29 Jul, in operation 1 Sep)IrelandNewVerification, certification, disclosure, prohibition
Regulation (EU) 2026/1744, Digital Omnibus on AI (OJ 24 Jul)EUNewRisk classification
EU AI Act, as amended (27 Jul)EURevisedRisk classification, prohibition, disclosure, competence
Article 50 transparency guidelines (Commission, final 20 Jul)EURevisedDisclosure
Code of Practice on AI-generated content (Commission opinion 8 Jul, AI Board endorsement 9 Jul)EURevisedDisclosure
Revised Ethics Guidelines on Generative AI (De Rebus, 1 Jul)South AfricaNew, draftVerification, competence, confidentiality

Six of the seven are in force; the South African draft has not been adopted by any regulator. The two documents that bind practitioners directly, the SRA notice and HC 142, both require verification of AI-assisted citations; HC 142 adds a declaration in affidavits and witness statements. The EU items move dates rather than duties: Annex III high-risk obligations now apply from 2 December 2027 and Annex I from 2 August 2028, while Article 50 transparency applied from 2 August 2026 as scheduled.

Method and caveats

An incident is a court, tribunal or disciplinary decision in which a party, lawyer, adjudicator or official put AI-generated material before a decision-maker and the decision records it: fabricated citations, invented quotations, doctored documents, or a ruling itself built on non-existent authority. The row date is the decision date. Rows are marked verified only when the primary document has been read; 13 of 122 are unverified as of 3 September 2026 and are shown in every headline count above, alongside the verified figure.

The tracker counts verified records, not every case worldwide. Charlotin’s database, the global baseline, held 2,009 entries on 3 September 2026 and logged 129 for July and 100 for August (August still back-filling), against the tracker’s 27 and 26. The tracker’s coverage outside the United States and United Kingdom depends on what secondary reporting surfaces, so the jurisdiction table measures where SafeLegalAI has verified rows, not where misuse is most frequent. Tool attribution is taken only from the record. Penalty sums exclude rows where the order gives no figure, and mix sanctions, fee-shifting and reimbursements. The actor field exists only for rows added or updated on 3 September 2026.

The dataset is exported at /tracker/incidents.json and the live figures are at /tracker/statistics. Data is licensed CC BY 4.0; cite safelegalai.com/tracker.

Three sentences journalists can quote

As of 3 September 2026, the SafeLegalAI Global Legal AI Incident Tracker records 54 verified or flagged incidents of AI misuse in legal proceedings dated between 1 July and 3 September 2026, 33 of them in United States courts.

Twenty of those 54 rows carry a referral to, or disposition by, a professional regulator, but only two regulator dispositions in the quarter are confirmed from the primary document: a public reprimand by the USPTO and a formal reprimand by a Dutch bar disciplinary council.

The largest stated penalty in the quarter is USD 46,511 in Kleyman Law Group v Kaloidis (Supreme Court of New York, Kings County, 4 August 2026), where 23 fictitious decisions were cited across 19 filings.

Sources

Primary documents for the five rows above, each re-read on 3 September 2026: