This week in legal AI safety: tracker reaches 122 incidents; SRA AI warning notice; Ireland's HC 142 live

This issue covers the relaunch week: 50 incidents added to the tracker from a July to September sweep, with rulings from England and Wales, the Netherlands, South Africa, New York, Illinois, New Mexico and the US Patent and Trademark Office, plus new rules from the Solicitors Regulation Authority, the Irish High Court, Connecticut and the European Union.

Incidents

In the High Court of England and Wales, Linden J found in Lodhia v Twelve Trees [2026] EWHC 1889 (KB) that a litigant in person had fabricated or doctored documents, including an altered court order and an N260 costs statement claiming about GBP 82,800, and rejected his explanation that they were AI hallucinations. The judgment of 24 July 2026 imposed indemnity costs and a three-year civil restraint order, referred the litigant to the Attorney General and referred his direct-access barrister to the Bar Standards Board. It is one of two UK rows this quarter; the other is the Crown Prosecution Service’s two non-existent extradition authorities in Tobosaru v Court of Law Craiova (approved judgment, Charlotin archive; Find Case Law).

The Raad van Discipline in ‘s-Hertogenbosch, a Dutch bar disciplinary council, reprimanded an advocate on 27 July 2026 for filing two tenancy defences with AI-generated ECLI citations, some non-existent and others pointing to unrelated divorce and criminal judgments. The council rejected a mere warning, found breaches of the core values of expertise and integrity and ordered EUR 1,250 in costs. The row is the tracker’s first from an EU member state (decision, tuchtrecht.overheid.nl).

Two US state courts set the quarter’s prices. In Kleyman Law Group v Kaloidis (Supreme Court of New York, Kings County, 4 August 2026) 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 per his testimony; the court imposed USD 46,511 (a USD 10,000 sanction plus USD 36,511.35 in fees and costs) and referred him to the Grievance Committee (order, Charlotin archive). In Scott v Illinois Human Rights Commission (Appellate Court of Illinois, 28 July 2026) the fine was USD 15,000 for ten false items, set at USD 1,500 each, with the opinion sent to the Attorney Registration and Disciplinary Commission (opinion PDF, Illinois Courts).

The US Patent and Trademark Office’s Office of Enrollment and Discipline publicly reprimanded patent attorney Brian Mitchell by settlement, final order 27 July 2026 and published 27 August 2026, after an AI-drafted claim-construction chart cited non-existent parts of the patent’s own specification and file history; IPWatchdog reported it as the USPTO’s first AI-predicated discipline order (final order, USPTO; row). The Supreme Court of New Mexico is reported to have held defence lawyer Stephen Aarons in contempt on 21 August, ordered USD 5,000 to the client protection fund and removed him from a murder appeal after a ChatGPT-assisted brief invented witnesses; the row is flagged unverified because only the Santa Fe New Mexican report (Yahoo mirror) has been located, not the order.

A full court of the High Court of South Africa, Gauteng Division, Johannesburg, upheld an acting judge’s findings in a parenting dispute on 31 July 2026 but identified at least 11 discrepancies in his judgment, among them a case that does not exist, Lubbe v Volkswagen SA, and a false S v Smith quotation. In FJL v TGO Opperman J referred the acting judge to the Legal Practice Council, reasoning that acting judges remain practising advocates subject to the Council rather than the Judicial Conduct Committee. No Council disposition is on record (judgment, Charlotin archive).

Regulator watch

The Solicitors Regulation Authority published a warning notice on misuse of AI on 17 August 2026, citing 42 reports of potential AI misuse received between July 2025 and July 2026. For solicitors in England and Wales it restates four things: they remain accountable for AI output; cited authorities must be genuine, relevant and verifiably cited; client data may enter an AI tool only with contractual, technical and organisational safeguards and never for model training without authority; and the compliance officer for legal practice must ensure supervision covers AI-assisted work. The notice creates no new rule, and failing to have proper regard to it, it says, puts a solicitor at risk of disciplinary action; our summer review reads it alongside the June 2026 AI sections of the effective-supervision guidance (notice, sra.org.uk; press release).

In Ireland, Practice Direction HC 142 came into operation on 1 September 2026, signed by the President of the High Court on 29 July. It applies to pleadings, submissions, affidavits, witness statements and expert reports in civil proceedings prepared after that date: every AI-assisted citation or proposition must be verified by a human, checking with another AI tool is stated to be insufficient, affidavits and witness statements must carry a declaration that generative AI was not used to generate their substantive content, and expert reports need the declaration or a disclosure. Sanctions listed are rejection of the document, adverse or personal costs, strike-out and referral to the regulator; Connecticut’s Practice Book Section 4-9, effective 23 June 2026, is the US counterpart, treating the filer’s signature as a certification that every AI-produced citation was independently verified (HC 142, courts.ie; Section 4-9, jud.ct.gov).

Regulation (EU) 2026/1744, adopted 8 July and in force across the European Union from 27 July 2026, defers the AI Act’s Annex III high-risk obligations to 2 December 2027 and the Annex I product-embedded obligations to 2 August 2028, so the date that would have bound Annex III point 8 systems (AI that assists a judicial authority in researching and interpreting facts and the law) from 2 August 2026 has moved by 16 months. Article 50 did not move. Its transparency duties and the Commission’s enforcement powers over general-purpose AI providers applied on 2 August 2026 as planned; the September EU AI Act update tabulates every date that moved and every one that did not (Official Journal text, EUR-Lex).

Vendor accountability

The Tools Directory launched this week with 128 external products. Forty-seven are verified, meaning at least four security facts are sourced to the vendor’s own trust, privacy or terms pages; the other 81 render noindex until enriched, and every field stays “unknown” until a source is found. Of the 128, 57 publish a commitment not to train models on customer data (38 of the 47 verified entries), 14 record an ISO 42001 certification and 9 carry accuracy evidence from someone other than the vendor, all counted from the directory YAML on 4 September 2026 and exported at /tools/tools.json.

Stat of the week

Lawyers get referred; litigants in person rarely do. Of the 54 tracker rows dated 1 July to 3 September 2026, 36 involve a lawyer and 9 a litigant in person; 18 of the 36 lawyer rows carry a regulatory referral or disposition, against 1 of the 9 litigant rows, and that one (Lodhia) is the referral of the litigant’s barrister.

Computed on 3 September 2026 from the actor and regulatoryOutcome fields of the incident YAML; the actor field was added this week and exists only on rows dated in the quarter, so no earlier comparison is possible. The Q3 incident report has the full tables and the statistics page recomputes the all-time figures on every build.

Tracker: 122 incidents worldwide across 15 jurisdictions, 50 added and 8 updated this week, 42 regulatory outcomes recorded, as of 3 September 2026.

SafeLegalAI is published by Cognesio LLP. LegalAI Space is an affiliated product (an Omnizio LP product built by the same team at Cognesio LLP) and is never scored on this site; see ownership and standards.