Mata v. Avianca: Why Lawyers Were Sanctioned for Submitting Fake Cases Generated by ChatGPT

On June 22, 2023, Judge P. Kevin Castel of the United States District Court for the Southern District of New York (S.D.N.Y.) issued the Opinion and Order on Sanctions in Mata v. Avianca, Inc. (22-cv-1461, 678 F.Supp.3d 443). In this case, the plaintiff's counsel cited in a brief fictitious cases found through research using ChatGPT and continued to insist on their existence even after the court and the opposing party raised questions, with the result that two attorneys and their law firm were ordered, among other things, to pay a monetary sanction under Rule 11 of the Federal Rules of Civil Procedure (FRCP). The point of rereading this case in 2026, when the use of generative AI has become established in legal work, is that the text of the order allows us to confirm that the reason for the sanctions was not the use of AI itself, but the failure to verify and the failure to correct after the problem was pointed out. This article was written after reviewing the original text of the order.

How the Litigation Began and the Fake Cases Created by ChatGPT

The case began as a personal injury lawsuit filed in New York state court on February 2, 2022 by the plaintiff, Roberto Mata, who alleged that he was injured when a serving cart struck his left knee on an Avianca flight. Avianca removed the case to federal district court on February 22 of the same year, asserting federal question jurisdiction under the Montreal Convention. Peter LoDuca of the law firm Levidow, Levidow & Oberman (the "Levidow firm"), counsel for the plaintiff, filed a notice of appearance as counsel in the district, but the attorney who actually handled the research and drafting was Steven Schwartz, who was not admitted to practice in that district.

On January 13, 2023, Avianca moved to dismiss on the ground that the two-year limitations period under the Montreal Convention had expired. In response, the Affirmation in Opposition filed by LoDuca on March 1 of the same year relied on cases that Schwartz had found using ChatGPT, but several of them (Varghese, Shaboon, Petersen, Martinez, Durden, Miller and others) did not exist. According to the text of the order, Schwartz repeatedly gave ChatGPT prompts to the effect of "show me cases supporting that the limitations period under the Montreal Convention was tolled by the defendant's bankruptcy," and ChatGPT fabricated fictitious cases each time. One of the fake cases listed the name of a real federal appellate judge as its author, and the order also directed the corrective measures described below on this point.

How They Persisted Even After Being Challenged

In its reply brief on March 15 of the same year, Avianca pointed out that it could not locate most of the cases cited in the affirmation. The court itself could not confirm the cases either, and in orders on April 11 and 12 of the same year it required LoDuca to submit copies of the opinions. The order states that, had LoDuca withdrawn the affirmation at this point and explained to the court the reality of Schwartz's research, the record in this case would have been quite different. However, neither attorney made a correction, and on April 25 of the same year LoDuca submitted excerpts of the fake cases in an affidavit (some were only excerpts rather than the full opinions, and for Zicherman it was stated that the case "could not be found").

On May 4 of the same year, the court first issued an Order to Show Cause (OSC) directing LoDuca to show why he should not be sanctioned under FRCP Rule 11(b)(2) and (c), 28 U.S.C. Section 1927, and the court's inherent power. The two attorneys first admitted that the cases did not exist in an affidavit dated May 25 of the same year, to which a screenshot of Schwartz's exchange with ChatGPT asking "Is Varghese a real case?" was also attached. On May 26 of the same year, the court issued an additional OSC extending the scope of potential sanctions to Schwartz and the Levidow firm, held a hearing on June 8 of the same year, and issued this sanctions order dated June 22 of the same year. The order treats not so much the carelessness at the time the affirmation was filed as the fact that the attorneys continued to maintain the fake cases without correcting them even after multiple challenges from the opposing party and the court as an important basis for finding subjective bad faith.

The Grounds and Content of the Sanctions Imposed by the Court

The order cited, among other things, that LoDuca signed and filed the affirmation without verifying a single case cited in it, and that Schwartz, while describing ChatGPT's research results as merely "supplementing" his research, later admitted that it was in fact his only research tool, and found that both attorneys violated Rule 11(b)(2) with subjective bad faith under Rule 11(c)(1). The court also held the Levidow firm jointly responsible, based on the provision in Rule 11(c)(1) that "absent exceptional circumstances, a law firm must be held jointly responsible." On the other hand, sanctions under 28 U.S.C. Section 1927 (a provision directed at abuse for the purpose of delaying proceedings) were not imposed, on the ground that the harm in this case was not delay of the proceedings as such.

The sanctions imposed were a USD 5,000 penalty (paid to the court) borne jointly and severally by the law firm and the two attorneys, plus the following ancillary measures. The attorneys were ordered to send the plaintiff, Mata, this order, the transcript of the hearing and the full set of the April 25 affidavit by registered mail; to likewise send this order and the relevant fake case to each real judge identified as the author of a fake case; and to file with the court, within 14 days, proof of these mailings. The court did not go so far as to order an apology, reasoning that a compelled apology is not a sincere apology. Also, because Avianca did not seek attorney's fees, there was no sanction on that point. As for the level of the monetary penalty, the order explains that it was limited to the extent necessary for deterrence, taking into account that the conduct was not motivated by financial gain or personal ill will, that the firm had already put in place training on the use of AI, and other factors.

"Using AI" Itself Was Not Treated as the Problem

At the outset, the order expressly states, to the effect that, just as with junior attorneys, a law firm's research support or the use of databases such as Westlaw and LexisNexis, there is nothing inherently improper about using a reliable generative AI tool to assist in advocacy. It then states that the "gatekeeping role to ensure the accuracy of filings" that FRCP Rule 11 imposes on attorneys is not lessened by the use of generative AI, and it gives the attorneys' abandonment of this role as the reason for the sanctions. In other words, what was sanctioned in this case was not the act of researching with generative AI itself, but the failure to personally verify the existence of the cases cited, and the maintaining of that falsehood even after repeated doubts were raised by the opposing party and the court, together with further misleading explanations in additional affidavits. Summarizing this as "using AI gets you sanctioned," while blurring this distinction, would misidentify where the duty to verify lies, as shown by this case.

Outcome on the Merits and How It Differs from the Sanctions Proceeding

As for the merits of this case, that is, the damages claim under the Montreal Convention itself, the publicly available court record confirms that a separate order dismissing the action on the ground that the limitations period had expired (Document 55 in the same case) was issued on June 22, 2023, the same date as this sanctions order. However, I have not been able to review the full text of that dismissal order on this occasion, and this article does not go into its content. The sanctions order is a separate proceeding based solely on the manner of counsel's filings, and it needs to be read separately from the merits.

The Practical Significance of Reading a 2023 Case in 2026

Although this case dates from 2023, fictitious case and statutory citations produced by generative AI have not decreased. As of the time of writing, it has been reported that many federal judges have issued standing orders requiring disclosure of the use of generative AI in filings or certification that a human has verified the AI-generated portions, and that the number of judges adopting such orders continues to grow. This structure does not become an issue in exactly the same form in Japanese corporate legal and litigation practice, but it is no reason to cite article numbers or court decisions found with generative AI without checking supporting materials. As a check you can carry out tomorrow, first review whether, for the briefs, opinions and internal legal opinions you have prepared, you have completed the work of matching each court decision and article number presented by generative AI, one by one, against primary sources (case law databases, the e-Gov law search and so on). Next, in my view, it is advisable to share as a written procedure an internal order of response under which, if the opposing party or a superior raises doubts about a source, you recheck it on the spot and do not keep maintaining the same argument while leaving the doubt unresolved. How to build a system in which a human makes the final check of generative AI output is explained in Why You Should Never Use Generative AI Answers As-Is and Always Have a Human Do the Final Check, and the specific steps for checking article numbers and sources yourself are explained in Always Verify Article Numbers and Sources Yourself: The Pitfalls of Generative AI. When reviewing the overall division of roles between generative AI and attorneys, see also What Is an AI Lawyer? How to Divide Work Between Attorneys and AI in Corporate Legal Practice in the Age of Generative AI.