This week in legal AI safety: SDT strike-off over AI citations; tracker at 633; three AI Act statutes filed

This issue covers 4 to 11 September 2026: the first Solicitors Disciplinary Tribunal case in England and Wales on a lawyer’s AI-generated citations, 509 rows added to the Incident Tracker including 476 US court decisions coded from primary documents, California’s SB 574 reaching the Governor, and the national EU AI Act statutes of Germany, Poland and Ireland entering the Regulation Tracker.

Incidents

In England and Wales, the Solicitors Disciplinary Tribunal struck Registered Foreign Lawyer Abhishek Kumar off the Register of Foreign Lawyers in a judgment dated 25 August 2026, after his 12 March 2026 answer to SRA proceedings and his 9 April explanatory email both relied on generative AI and both contained citations that did not exist, did not contain the quoted passages or did not support the propositions cited. The tribunal found breaches of Principles 1, 2 and 5 and Code paragraph 1.4, made no order as to costs, and its case page records this as the first time a lawyer’s use of AI in legal proceedings had been litigated before it. Our brief and the tracker row carry the detail (judgment PDF, solicitorstribunal.org.uk).

In California, the Appellate Division of the Orange County Superior Court sanctioned two attorneys from the Law Office of Vincent W. Davis & Associates, USD 2,500 for supervising counsel and USD 1,500 for the associate, in an order filed 4 August 2026 and modified on 5 August, after a supersedeas petition cited four suspicious authorities and the associate admitted using ChatGPT and Google Gemini; the court directed the clerk to serve the order on the State Bar and both attorneys to report it personally. The certified-for-publication order in Southland Homes v. Lam, posted by California Courts on 10 September, is one of this week’s seven new US rows (order PDF, courts.ca.gov).

Two state appellate courts priced unverified AI research on 8 September 2026. The Missouri Court of Appeals, Eastern District, denied a mandamus petition and ordered USD 1,750 in attorney-fee sanctions under Rule 55.03 after relator’s counsel admitted two non-existent cases came from generative AI (State ex rel. Schnell v. Richardson; opinion PDF mirrored by Damien Charlotin; the Missouri Courts page was not reachable to our fact-checker), and the Minnesota Court of Appeals affirmed dismissal and USD 2,341.25 in fee sanctions against a self-represented litigant who admitted using AI without checking the citations (Asiedu v. Eastlund; opinion PDF, Minnesota State Law Library).

The tracker reached 633 rows this week, 509 of them added since 4 September 2026, because 476 US court decisions from the us-genai-court-opinions dataset were written back as status: provisional records on 7 September, each linked to its court document but not yet re-read by an editor; they render noindex and queue at /tracker/provisional. The dataset’s own changelog records that on 8 September 2026 an independent re-read of a 50-row random sample found 22 rows (44 percent) with at least one coding-column error and no passage or identification errors, that every row was then re-read from the court document, and that a 95-row slice coded twice reached 89 percent agreement on outcome (README, github.com/SafeLegalAI/us-genai-court-opinions).

Regulator watch

In California, SB 574 was “Enrolled and presented to the Governor at 2 p.m.” on 9 September 2026, according to the Legislature’s bill history. The bill would add Business and Professions Code section 6068.1, barring attorneys from delegating the practice of law to generative AI and requiring verification of every AI-produced citation, and would amend Code of Civil Procedure section 128.7 so that no filed paper may contain a citation the responsible attorney has not personally verified; under article IV, section 10(b)(2) of the state constitution the Governor has until 30 September 2026 to return it. Our 7 September brief now carries a dated update (bill history, leginfo.legislature.ca.gov).

Three national statutes implementing the EU AI Act entered the Regulation Tracker on 11 September 2026, each read from the official text; two are in force and one awaits commencement. Germany’s KI-MIG, in force from 29 July 2026, makes the Bundesnetzagentur the default market-surveillance authority and creates an independent chamber for Annex III points 1, 6, 7 and 8, which include AI used to assist judicial authorities (gesetze-im-internet.de). Poland’s act of 3 July 2026 on AI systems, in force from 11 August 2026, establishes the Komisja Rozwoju i Bezpieczeństwa Sztucznej Inteligencji as supervisor and single point of contact (eli.gov.pl), and Ireland’s Regulation of Artificial Intelligence Act 2026, enacted on 21 July 2026 with commencement by ministerial order, provides for Oifig IS na hÉireann, a statutory AI register and a complaint route (enterprise.gov.ie).

Thirteen of the tracker’s most-linked documents were re-opened against their official pages on 11 September 2026 and none had changed: the SRA’s misuse-of-AI warning notice and compliance tips, the judiciary.uk AI guidance of October 2025, the Bar Standards Board guidance of May 2026, Ireland’s Practice Direction HC 142, NSW Practice Note SC Gen 23, the Federal Court of Australia’s GPN-AI, the New Zealand courts’ guidelines, Singapore’s Registrar’s Circular 1 of 2024, the Ontario Superior Court practice directions, the Federal Court of Canada notice, and the consolidated EU AI Act and Regulation (EU) 2026/1744. Each record now shows that date under lastVerified (SRA warning notice; notice, sra.org.uk).

Vendor accountability

The weekly re-verification of ten Tools Directory records on 11 September 2026 moved three to verified, Casepoint, Thomson Reuters CoCounsel (the former Casetext product) and Caseway, and one the other way: Clio Duo, where no current Clio page could be found to support the recorded private-deployment position, left the record with three sourced security facts against the four the directory requires. Casepoint’s product page now states that “IQ Assist helps your team work faster with privately deployed LLMs” (casepoint.com; directory record).

Two general-purpose enterprise assistants used in firms lost a retention claim. ChatGPT Enterprise moved from zero retention “yes” to “no” because OpenAI’s data-governance page describes a custom retention policy “with a minimum of 90 days” for workspaces, with the API’s zero-data-retention option a separate product (OpenAI Academy), and Claude for Enterprise now carries the caveat that prompts and outputs sent to Anthropic’s covered models “are retained for 30 days” regardless of a zero-data-retention agreement, with published data residency narrowed to the United States (support.claude.com).

Stat of the week

Of the 633 incident rows on the tracker at 11 September 2026, 146 are verified (an editor or the sweep read the primary document), 476 are provisional (coded from the court document by the dataset pipeline and awaiting an editor) and 11 are unverified (press-reported only). Fifty rows record a regulator’s or disciplinary body’s action.

Computed on 11 September 2026 with grep -h '^status:' src/content/incidents/*.yaml | sort | uniq -c and grep -l '^regulatoryOutcome:' over the same folder; the provisional tier exists only since 7 September 2026, so there is no earlier comparison. The sanctions ledger and regulator outcomes report hold the verified-tier figures as of 5 September.

Tracker: 633 incidents worldwide across 15 jurisdictions, 509 added and 126 updated this week, 50 regulatory outcomes recorded, as of 11 September 2026.

SafeLegalAI is published by Cognesio LLP. LegalAI Space is an affiliated product (a Cognesio LLP product) and is never scored on this site; see ownership and standards.