A federal court fined a public defender $1,500 for a fake citation, then declined to say AI wrote it

The problem

An AI-fabricated citation and an AI-caused sanction are not the same fact, and cases like this one are routinely listed as the second when the record only supports the first. A brief can carry a citation with every hallmark of a machine-generated case while the court expressly declines to find that any AI tool produced it. The honest reading is what the order says, not what the shape of the error suggests.

What was built

Nothing was built, and that is the point of this entry. On 21 November 2024, Assistant Federal Defender Andrew Francisco filed a motion to unseal in a federal criminal prosecution in Sacramento, whose principal authority was United States v. Harris, 761 F. Supp. 409, 414 (D.D.C. 1991), quoted for the proposition that “The legislative history indicates that Congress intended to shield the names and addresses of child victims and witnesses from public disclosure” (source). The prosecution checked it: the citation 761 F. Supp. 409 belongs to Harris v. Murray, a 1990 prisoner civil-rights decision from the Eastern District of Virginia, and the quoted sentence returned zero results on Westlaw and Lexis (source). In the court’s words, “‘United States v. Harris, 761 F. Supp. 409, 414 (D.D.C. 1991)’ is not a real case” (source).

The reply brief did not concede it; it said the quote actually came from United States v. Broussard, where the sentence also does not appear, and at a 9 December 2024 hearing, when Francisco could not find the quotation in the Broussard opinion the court handed him, he suggested it came from a second Broussard opinion the court had already read and confirmed did not contain it (source). The court found the citation had the shape of a machine-generated one: “The citation has all the markings of a hallucinated case created by generative artificial intelligence (AI) tools such as ChatGPT and Google Bard” (source). And then it declined to decide the question: “The Court need not make any finding as to whether Mr. Francisco actually used generative AI to draft any portion of his motion and reply, including the fictitious case and quotation” (source).

The attorney’s own account, quoted in the order, is a denial: he states he “did not use and ha[s] never used AI (artificial intelligence) to draft any of my motions,” says he drafted the passage “hastily,” and calls the fake case “an inadvertent drafting error I cannot begin to explain” (source). The court found that response “inadequate and not credible,” but its finding of non-credibility is about candour, not about AI (source). So the record has an AI-shaped hole in it and no one filled it: no tool, no vendor, no prompt, no admission.

The outcome

The sanction: $1,500, and not under Rule 11. The order issued on 17 January 2025 under Eastern District of California Local Rule 180(e) and the court’s inherent authority, because Rule 11 does not reach a criminal case, and it holds that “Assistant Federal Defender Andrew Francisco, defense counsel, is personally sanctioned in the amount of $1,500” (source). The court concluded that sum was “proportionate to his conduct” (source).

The finding was of bad faith, not carelessness. The court held Francisco’s “inaccurate and misleading statements were not inadvertent as claimed, but knowing and made in bad faith” (source).

The consequences ran past the money. The Clerk was ordered to serve the order on the District of Columbia Bar, “of which Mr. Francisco is a member (DC Bar No. 1619332),” and the State Bar of California, and on every district and magistrate judge in the district, because “this is the first time it has arisen in this federal district, raising the importance of educating and deterring the larger bar from repeating similar conduct” (source).

It was collected. The docket records receipt number 200011278 for $1,500.00 in sanctions from Andrew Francisco on 6 February 2025, and reconsideration was denied on 9 April 2025 (source).

What is not settled is whether it stands. Counsel appealed to a district judge and the appeal was severed into In Re: Appeal of Andrew Francisco, No. 2:25-mc-00112 (E.D. Cal.); at a 7 July 2025 hearing before Judge Dale A. Drozd, counsel “requested court vacate sanction order and remand matter to different magistrate judge,” and the matter was taken under submission (source). As of 2026-08-16 the public docket shows no order resolving the appeal: its two most recent entries, dated 1 November 2025 and 3 August 2026, are notice filings, not a ruling, so we do not report the sanction as final (source).

How this was verified

This is a court-origin record, not a vendor claim, and it carries a green badge under TIN’s public-record standard. On 2026-08-16 every figure and quotation above was re-checked verbatim against the E.D. Cal. sanction order (signed 2025-01-17, No. 2:24-cr-0280-DJC) as reproduced live on FindLaw and cross-read against the Leagle copy, and the payment and appeal facts were re-checked the same day on the live CourtListener docket for No. 2:25-mc-00112. The characterisation of the citation as AI-hallucinated is independently reported by Eugene Volokh (Reason / The Volokh Conspiracy, 2025-01-23).

Two honest limits sit on this page and neither is a sourcing gap. First, no source establishes that AI caused this: the court refused to find it and the attorney denies it, so anyone who lists this as an “AI sanction” has added a fact the record does not contain. Second, the sanction is not final: the de novo appeal is undecided on the public docket we can reach, and a ruling that vacates it would change this story. No confirmation was sought from Francisco, his office, or any vendor: green never depends on the subject confirming anything.

Related case files

The mirror image is Gauthier v. Goodyear, where a federal court did attribute the fabricated citations to generative AI and ordered a $2,000 sanction plus a generative-AI CLE; reading the two together shows exactly what a court has to say for an “AI sanction” label to be earned. For the same failure inside a large firm, Wadsworth v. Walmart shows a judge fining Morgan and Morgan lawyers $5,000 after AI hallucinated eight of nine cited cases, with the AI use admitted rather than denied. And in another federal criminal matter, United States v. McGee, a court reached a $5,000 sanction over a ghostwritten brief built on fabricated citations, the criminal-docket counterpart to Hayes.

Sources

Tier 1 = the primary court order and docket. Tier 2 = independent press naming the parties. Each source was checked live on 2026-08-16.

1. FindLaw, “United States v. Daragh Finbar Hayes” (E.D. Cal. sanction order, Hon. Chi Soo Kim, signed 2025-01-17) (Tier 1, verbatim reproduction of the signed order; every quotation above checked against it live). https://caselaw.findlaw.com/court/us-dis-crt-e-d-cal/116862866.html
2. CourtListener (Free Law Project), “In Re: Appeal of Andrew Francisco, No. 2:25-mc-00112 (E.D. Cal.)” (Tier 1, official docket; source for the 2025-02-06 payment receipt, the 2025-04-09 reconsideration denial, and the undecided appeal status through the 2026-08-03 entry). https://www.courtlistener.com/docket/69878769/in-re-appeal-of-andrew-francisco/
3. Reason / The Volokh Conspiracy, “Federal Public Defender Submits Brief with Nonexistent Citation, Apparently Refuses to Admit This to the Judge at a Hearing” (Eugene Volokh), 2025-01-23 (Tier 2, independent secondary; corroborates the fabricated citation and the hearing, does not establish AI causation). https://reason.com/volokh/2025/01/23/federal-public-defender-submits-brief-with-nonexistent-citation-apparently-refuses-to-admit-this-to-the-judge-at-a-hearing/