newsletterAI in CourtsAI HallucinationsAI GovernanceVendor Watch

This week in legal AI safety: California's top court refers AI misquotes; SB 574 signed

22 September to 5 October: a California Supreme Court referral, SB 574 enacted, two English court guides on AI, and Harvey named in a New York filing.

Harvey

Edited and verified by SafeLegalAI

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

This issue covers 22 September to 5 October 2026: nine incident records added on 5 October, including rulings from California, Massachusetts and Canada, California’s SB 574, two court guides for England and Wales, and vendor disclosures involving Harvey and Juro.

Incidents

California’s Supreme Court refers an AI-affected misquote to the State Bar

The Supreme Court of California referred an AI-affected briefing error to the State Bar of California on 24 September, in footnote 7 of its Cervantes v. Bianco opinion (S295866). Lawyers for Riverside County Sheriff Chad Bianco had challenged standing with misquoted passages from, and an incorrect recitation of the holding in, the court’s own 1989 Common Cause decision; the responsible firm apologised and withdrew the argument. Our incident record names no firm, because the opinion lists two for the sheriff and the footnote does not say which was responsible.

A Massachusetts court revokes pro hac vice admission over fictitious citations

Judge Angel Kelley of the US District Court for the District of Massachusetts revoked a California-barred lawyer’s pro hac vice admission on 25 September in Aguilar v. The Crawford Group and ordered his firm to pay the Enterprise defendants’ fees and costs, capped at $10,000, under Rule 11 (order, Dkt. 155, p. 18). Counsel had said he believed the enterprise-level AI version he used did not hallucinate cases; the court found citation errors across three briefs filed over months. The exact amount is still open, and the incident record will add it once fixed.

Canada’s Federal Court treats fake citations as an abuse of process

Justice Lafrenière of the Federal Court of Canada refused the applicants an extension of time and dismissed their refugee leave application on 28 September in Ahmed v. Canada (Citizenship and Immigration), 2026 FC 1202, after their proposed memorandum listed five cases that could not be found at the citations given (reasons, paras 11–15). He wrote that citing fake cases “is an abuse of process and is tantamount to making a false statement to the Court” and, with no reply from the applicants, concluded that AI was used without careful human judgment. No costs were ordered, because the respondent sought none (incident record).

Regulator watch

England and Wales: the TCC Guide resolves doubt in favour of disclosure

The Technology and Construction Court Guide 2026, published on 1 July, says at para 1.6.2 that whether legal representatives disclose AI use depends on context, but “any doubt as to whether the fact that AI has been used should be mentioned must be resolved in favour of disclosing such use” (guide, section 1.6). Para 1.6.4 requires representatives to independently verify AI-assisted research and case citations and to be in a position to confirm they have done so. The regulation record was added on 5 October; the guide itself dates from July.

The Administrative Court lists sanctions up to police referral

The Administrative Court Judicial Review Guide 2026, dated August 2026, says legal representatives and litigants in person should take reasonable care when using AI tools to draft pleadings, witness statements and skeleton arguments, in particular over citations, legal propositions and facts (guide, para 2.3.1). Under para 2.3.2, failure may result in costs sanctions, strike-out, referral to a professional regulator, contempt proceedings or referral to the police, citing R (Ayinde) v Haringey [2025] EWHC 1383 (Admin). The regulation record also covers para 20.2.5 on checking authorities.

California’s SB 574 becomes Chapter 858

California’s governor approved SB 574 on 30 September, and it was chaptered the same day as Chapter 858, Statutes of 2026, according to the official bill history. From 1 January 2027, the date it takes effect under article IV, section 8(c)(1) of the California Constitution because it has no urgency clause, it bars attorneys from delegating the practice of law to generative AI, and a new Code of Civil Procedure §128.7(b)(2)(A) says filed papers must not contain citations the responsible attorney has not personally verified. Our news brief and regulation record set out the arbitrator provision and how renumbering changes the target of the monetary-sanctions bar in §128.7(d)(1).

Vendor accountability

Counsel names Harvey after a New York “wakeup call”

Defence counsel in Hill v. Foundation Media told the US District Court for the Southern District of New York on 1 October that “the original version of Harvey, as potentially supplemented by one or more updates referred to collectively as Harvey 2.0, was used” to prepare a response that the court’s 22 September order (Dkt. 54) had questioned over quotations attributed to Piazza and a citation to Zhao (letter, Dkt. 58). Judge Arun Subramanian’s 29 September order took no further action, called the episode “a wakeup call” and required a letter disclosing the brand and version of AI used. Our incident record records counsel’s disclosure, not a finding that Harvey caused the errors; as of 5 October 2026 it is the only tracker record naming Harvey as the tool used.

Juro’s privacy policy lists Claude for an optional contract service

Juro’s privacy policy, last updated on 29 September 2026, lists an enterprise version of Anthropic’s Claude as an internal productivity tool and says Juro may use Claude to help perform its optional AI Managed Service on customer contracts, with processing in the US, UK and EEA and, for some internal processes, other countries. The Juro directory entry, re-checked on 5 October, records EEA hosting for platform contract data and was promoted to verified that day with Jus AI and Lawpath, bringing verified directory entries to 62. The policy is the vendor’s own statement.

Stat of the week

106 of the 162 verified incident records (65%) carry a 2026 date, counted from each record’s date field as of 5 October 2026. Seventeen of those are dated September and two October.

Four other records added on 5 October: Chirece C. v. Commissioner of Social Security, a $500 fine in Michigan federal court (record); Zhejiang Dingli v. United States, a pending show-cause order in the Court of International Trade (record); Shannon v. Shannon, where the Nebraska Supreme Court struck a self-represented litigant’s brief (record); and R. M. v. Mohr, a $250 sanction on a self-represented appellant in the Oregon Court of Appeals (record).

Additions and updates count public incident files added or modified between 22 September and 5 October 2026: Posey v. Bushnell on 22 September and nine records on 5 October, plus six re-checked records. Jurisdictions are counted by each record’s jurisdiction tag, and the “other” tag counts once. The outcome count measures populated regulatory-outcome fields, which include pending referrals such as Cervantes.

Tracker: 170 public incident records worldwide across 15 jurisdictions, 10 added and 6 updated between 22 September and 5 October, 51 regulatory-outcome fields populated, as of 5 October 2026.

Sources

External sources retrieved 5 October 2026:

Internal references: Cervantes incident; Aguilar incident; Ahmed incident; TCC Guide regulation; Administrative Court guide regulation; SB 574 regulation; SB 574 news brief; Hill incident; Juro directory entry; public tracker.

Tracker figures are counted from the incident files in the working dataset on 5 October 2026; 476 provisional rows are excluded from the public count.