AI-Fabricated Citations: The Court Sanctions Orders, Dated

Courts have not sanctioned lawyers for using a language model. In every published order so far they have sanctioned lawyers for signing a filing whose authorities they had not read, and for what they did after the problem was pointed out.

Mata v. Avianca, and what it decided

The anchor case is Mata v. Avianca, Inc., No. 22-cv-1461 (PKC) in the Southern District of New York. On 22 June 2023 Judge P. Kevin Castel issued an opinion and order imposing a joint and several penalty of $5,000 on two attorneys, Steven A. Schwartz and Peter LoDuca, and their firm, and directing them to send the opinion to the judges whose names had been attached to the invented decisions. The underlying filings cited authorities including a purported “Varghese v. China Southern Airlines” that does not exist; the text had come from ChatGPT, which also confirmed on request that the cases were real.

The order is worth reading rather than summarising because of what it declines to say. Judge Castel wrote that there is nothing inherently improper about using a reliable artificial intelligence tool for assistance, and located the wrong in the abandonment of responsibility — the failure to verify, and the continued defence of the fabricated cases after opposing counsel and the court had flagged them. You can read the docket through govinfo’s United States Courts Opinions collection, which carries the published opinions in the case for free.

The orders, with their dispositions

These are the ones with a written disposition, in date order. Amounts and terms should be confirmed against the docket before you rely on them; several were modified on reconsideration.

• Mata v. Avianca (S.D.N.Y., 22 June 2023) — $5,000 joint penalty against two attorneys and the firm, plus notice to the misattributed judges. No referral to the state bar in the order itself.
• Park v. Kim, 91 F.4th 610 (2d Cir., 30 January 2024) — the Second Circuit referred the attorney to its own grievance panel after a reply brief cited a non-existent decision produced by ChatGPT. The referral, not a fine, was the sanction; the appeal itself was decided on other grounds.
• Kruse v. Karlen, 692 S.W.3d 43 (Mo. Ct. App., 13 February 2024) — a Missouri appellate court dismissed the appeal outright and awarded damages for a frivolous appeal, in a brief where the great majority of the cited authorities were fictitious and the appellant was self-represented but had paid an online “AI consultant”.
• Wadsworth v. Walmart Inc. (D. Wyo., February 2025) — Judge Kelly H. Rankin sanctioned attorneys at a large plaintiffs’ firm over motions in limine containing invented citations, including revocation of one attorney’s pro hac vice admission alongside monetary sanctions. Read the order for the per-attorney amounts.
• Lacey v. State Farm General Ins. Co. (C.D. Cal., May 2025) — Special Master Michael Wilner ordered the plaintiff’s firms to pay the costs of the exercise after a brief containing fabricated authorities, and struck the offending filings. Notably the problem survived a round of internal review before filing.
• Coomer v. Lindell (D. Colo., July 2025) — Judge Nina Y. Wang sanctioned two attorneys individually for a filing with numerous defective citations, rejecting the explanation that an unchecked draft had been filed by mistake.

Which power the court used

The orders do not all rest on the same authority, and the difference matters if you are predicting exposure. In federal civil practice the usual hook is Federal Rule of Civil Procedure 11, under which signing a paper certifies that the legal contentions are warranted by existing law after an inquiry reasonable under the circumstances. Rule 11(c) sanctions are discretionary and are supposed to be limited to what suffices to deter repetition. Separately, 28 U.S.C. § 1927 reaches an attorney who multiplies proceedings unreasonably and vexatiously, and courts also have inherent authority, which generally requires a finding of bad faith. Appellate courts and state courts have their own frivolous-appeal rules, which is how Kruse produced an award of damages rather than a Rule 11 penalty.

Professional-conduct exposure is a separate track from the court’s own sanction. The American Bar Association addressed generative AI directly in Formal Opinion 512, issued 29 July 2024, framing the duties under Model Rules 1.1 (competence), 1.6 (confidentiality), 3.3 (candour to the tribunal) and 5.1/5.3 (supervision). Individual jurisdictions have gone further; see the state bar ethics opinions for the ones that impose disclosure or verification duties by name.

The conduct being punished is not the hallucination

Read the six orders together and the fabrication is never the operative finding. Three things recur. First, the citation was not checked against a reporter or a database before filing, so the certification in the signature block was false when it was made. Second, when challenged, the filer defended the citation instead of withdrawing it — in Mata the attorneys produced excerpts of the non-existent opinions on the court’s order rather than admitting the cases did not exist, and that escalation is what the opinion dwells on. Third, in the firm cases there was a supervision gap: the drafting was delegated, the review was nominal, and nobody in the chain treated an unfamiliar citation as something to open.

The mechanism underneath is unremarkable to anyone who has read how these models actually produce text. A model returns a high-probability continuation, and a citation — volume, reporter, page, year, parenthetical — has an extremely regular surface form. Producing a plausible one is exactly what the model is good at; producing a true one requires that the specific case be represented well enough to be recalled, which is a different property and is not guaranteed by fluency.

What the orders imply about checking

None of the orders requires a particular tool or a disclosure. What they converge on is a verification step that is independent of the system that produced the text: open the cited authority in a citator or a primary database, confirm that the case exists, that the proposition is in it, and that it has not been reversed. A retrieval-augmented research product reduces the invention rate but does not remove the step, because a real case can still be cited for a proposition it does not contain — and that failure is invisible to a check that only asks “does this case exist”.

For an engineering team building anything that emits citations, the practical lesson is that the check has to be a separate pass with its own failure mode, not a prompt instruction. Asking the model to confirm its own citations reproduces the Mata failure exactly: that is precisely what the attorney did, and the model said yes.