A Connecticut federal judge fines a lawyer $3,500 and refers him for discipline over AI-fabricated legal authorities

Verification status: IN CHECKING — not publish-ready, not pending, not verified.

The problem

Barteca Holdings, which operates the bartaco restaurants, sued the operators of two tacobarn restaurants for trademark and trade-dress infringement in the U.S. District Court for the District of Connecticut, where defense counsel Hilary Miller filed a Motion to Dismiss (ECF 24) and a Motion to Exclude Extrinsic Evidence (ECF 28) (source). Reviewing those motions, the court, in its own words, “discovered numerous case quotes that do not exist, erroneous citations, and misrepresentations of the law” (source).

What was built

No system was built or deployed — this is a court-origin AI-governance honest-negative, a federal court’s response to unverified generative-AI output submitted as law (source). On the record, Miller disclosed that “he used a tool called Open Law, which he paid for, for suggestions on arguments and additional cases. He also used Claude and Chat GPT for additional suggestions” (source). The court framed the matter as counsel’s “submission of unverified AI-generated legal authorities” and treated the resulting citations as a Rule 11 problem in their own right (source).

The outcome

After an Order to Show Cause (ECF 33) and a show-cause hearing, Judge Vernon D. Oliver issued the Order Imposing Sanctions (ECF 43) on August 4, 2026 (source). The court found that “Attorney Miller acted with subjective bad faith in submitting filings containing legal authorities that he had not reviewed or verified,” and concluded that his “submission of unverified AI-generated legal authorities warrants sanctions” (source). It held that “a monetary sanction of $3,500, together with a referral of this matter to the Grievance Committee pursuant to Local Rule 83(c)(2), is appropriate,” and directed that “Attorney Miller should deposit the monetary sanction with the Clerk of Court on or before August 14, 2026” (source). An independent legal-commentary post by Michael Berman for eDiscovery LLC / EDRM, published Aug. 17, 2026, reports the same outcome, writing that “the court imposed a sanction of $3,500 ‘together with a referral of this matter to the Grievance Committee….’” (source). The American Bar Association’s flagship newsroom, the ABA Journal, reported the sanction as its own news on Aug. 6, 2026, writing that “a Connecticut federal judge has imposed a $3,500 fine on a defense attorney who submitted two motions containing artificial intelligence-generated mistakes,” that the judge “referred him to the Federal Grievance Committee for potential disciplinary action,” and that Oliver “found that although Miller expressed remorse about the AI errors and promised to make changes to his verification process, he still showed ‘subjective bad faith’” (source).

Conflation guards

• The $3,500 and the Grievance referral are the Rule 11 sanction — the underlying dispute is a trademark / trade-dress case (bartaco vs. tacobarn), and the sanction is not a merits award on the trademark claims (source).
• Three AI tools are named (Open Law, Claude, ChatGPT), and the court did not pin the hallucinations to a single product, so coverage that flatly says “ChatGPT” alone would be incomplete (source).

How weak is the weakest source

The headline figures rest first on a Tier-1 primary — the court’s own Order Imposing Sanctions (ECF 43) — now joined by two Tier-2 independents: an eDiscovery LLC / EDRM commentary post by Michael Berman (Aug. 17, 2026) and the ABA Journal’s news report (Kevin Davis, Aug. 6, 2026), each of which states the $3,500 sanction and the Grievance referral. The weakest load-bearing source here is the Berman post: it is a lawyer’s legal-commentary blog, not the primary filing and not a wire newsroom, so it corroborates the $3,500 and the Grievance referral but adds no independent measurement of its own (source). The ABA Journal is the stronger of the two secondaries — a national bar-association newsroom with a named byline — but it credits the trademark-merits background to Law360 and reports the sanction facts as its own; the court order remains the substance in either case. Law360 Pulse also carried the $3,500 but sits behind a paywall and is not archive-bindable, so it is not counted. Corroboration is two_independent.

How this was verified

Method — court-adjudicated, retrieved live and byte-identified. Every figure is quoted verbatim from the Order Imposing Sanctions (ECF 43, filed Aug. 4, 2026, Judge Vernon D. Oliver), retrieved from the CourtListener RECAP copy and byte-identified against it (local sha1-b32 2KWNS2SXVGJX57RKNQV2YAL6WDED3HUQ, 280,345 B) (source). No Wayback capture of the order exists yet because it was filed the same day, so RECAP is the durable primary host. Two independent secondaries corroborate the sanction facts and are each Wayback-archived and byte-tied: the eDiscovery LLC / Berman post (capture 20260821203822) and the ABA Journal news report (capture 20260821211252). Verification date: 2026-08-21.

Path to green

Every figure is established by the court’s own filing and needs no client confirmation. The open items are docket-only and asserted by no claim: whether Miller deposits the $3,500 by the Aug. 14, 2026 deadline, the outcome of the Grievance Committee referral, and any appeal or reconsideration. The $3,500 and the Grievance referral are now corroborated by two free, archive-bound independents — the eDiscovery LLC / Berman post (Wayback capture 20260821203822) and the ABA Journal news report (Wayback capture 20260821211252), both byte-tied — which lifts corroboration to two_independent. That is the strongest lift the independent public record is presently expected to yield; further corroboration would come only from additional free, archive-bindable newsrooms carrying both figures (Law360 Pulse carries the $3,500 but is paywalled and not archive-bindable). No confirmation from the court or counsel is sought or needed — the sanction and quotes are already verifiable from the primary.

Related case files

• Another 2026 Rule 11 sanction where counsel named both Claude and ChatGPT as the source of the fabricated citations
• A Connecticut sanction on the same AI-hallucinated-citation fact pattern, decided at the state’s high court
• The costliest US AI-hallucination sanction on record, showing the far end of the monetary range this $3,500 sits within

Sources

1. U.S. District Court for the District of Connecticut · Barteca Holdings LLC v. tacobarn Newtown LLC — Order Imposing Sanctions (ECF 43) · 2026-08-04 · Tier 1 (primary, court-adjudicated; first-party filing) · https://storage.courtlistener.com/recap/gov.uscourts.ctd.169303/gov.uscourts.ctd.169303.43.0.pdf
2. eDiscovery LLC (Michael Berman, EDRM) · Cite Checking to Find Hallucinated Cases Deemed Insufficient – Part II: Verification and Candor Are Expected · 2026-08-17 · Tier 2 (independent legal-commentary secondary; quotes the order verbatim, independent of court, counsel and parties) · archived https://web.archive.org/web/20260821203822/https://www.ediscoveryllc.com/cite-checking-to-find-hallucinated-cases-deemed-insufficient-part-ii-verification-and-candor-are-expected/ · https://www.ediscoveryllc.com/cite-checking-to-find-hallucinated-cases-deemed-insufficient-part-ii-verification-and-candor-are-expected/
3. ABA Journal (Kevin Davis) · Federal judge fines attorney for AI mistakes in taco biz beef · 2026-08-06 · Tier 2 (independent national bar-association newsroom, named byline; reports the $3,500 sanction, the Federal Grievance Committee referral and the “subjective bad faith” finding as its own news, independent of court, counsel and parties) · archived https://web.archive.org/web/20260821211252/https://www.abajournal.com/news/article/federal-judge-fines-attorney-for-ai-mistakes-in-taco-biz-beef · https://www.abajournal.com/news/article/federal-judge-fines-attorney-for-ai-mistakes-in-taco-biz-beef