Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023), is the sanctions case that made "ChatGPT invented my citations" a known risk in every law office. On June 22, 2023, Judge P. Kevin Castel fined two lawyers and their firm $5,000 and ordered letters to their client and to the judges whose names were put on fake opinions. The court did not punish the use of AI. It punished what the lawyers did after they were told the cases could not be found.
The opinion is more useful than the headline. It is a set of findings about who signed what, who read what, and when each lawyer knew enough to stop. This post walks through it from the sanctions opinion itself. For the cases that followed, see our running list of AI hallucination sanctions cases.
The underlying case
Roberto Mata alleged that a metal serving cart struck his left knee on an Avianca flight from El Salvador to John F. Kennedy Airport in August 2019. He sued in New York state court on February 2, 2022, and Avianca removed the case to federal court, asserting jurisdiction under the Montreal Convention.
Steven A. Schwartz of Levidow, Levidow & Oberman had filed the state complaint. He was not admitted in the Southern District of New York, so his colleague Peter LoDuca filed the federal notice of appearance while Schwartz kept doing the substantive work.
On January 13, 2023, Avianca moved to dismiss. Its argument was simple: the Montreal Convention sets a two-year limit, and Mata filed too late. Mata's answer was that Avianca's bankruptcy had tolled the limit. That argument needed case law, and the real case law was against it.
The timeline that led to sanctions
| Date (2023) | What happened |
|---|---|
| March 1 | LoDuca files an "Affirmation in Opposition," written by Schwartz, citing and quoting decisions said to be in the Federal Reporter, Federal Supplement and Westlaw |
| March 15 | Avianca's reply says its counsel could not locate most of the cited cases, and that the few it found did not support the propositions cited |
| April 11 and 12 | The court orders LoDuca to file copies of eight cited decisions, then a ninth |
| April | LoDuca seeks an extension, telling the court he is out of the office on vacation |
| April 25 | LoDuca files an affidavit attaching what it calls copies or excerpts of all but one case, and says he could not locate Zicherman v. Korean Air Lines |
| May 4 | Order to show cause why LoDuca should not be sanctioned under Rule 11, 28 U.S.C. § 1927 and the court's inherent power |
| May 25 | Schwartz's affidavit is the first acknowledgment by any respondent that the cases came from ChatGPT |
| May 26 | Supplemental order to show cause directed to Schwartz and the firm, including the notarization of the April 25 affidavit |
| June 8 | Sanctions hearing, with both lawyers questioned under oath |
| June 22 | Sanctions opinion, and a separate opinion granting the motion to dismiss |
At no point, the court found, did any respondent write to withdraw the March 1 filing.
What the fake cases looked like
The respondents ultimately admitted that six decisions were generated by ChatGPT and do not exist, among them Varghese v. China Southern Airlines, Shaboon v. Egyptair, Petersen v. Iran Air, Martinez v. Delta Airlines, Estate of Durden v. KLM Royal Dutch Airlines and Miller v. United Airlines.
The court's findings on "Varghese" are a field guide to what a fabricated opinion looks like:
- It was presented as an Eleventh Circuit decision, but the Clerk of that court confirmed it was not authentic.
- Its docket number belonged to a different case, and its Federal Reporter citation pointed to J.D. v. Azar, a D.C. Circuit decision.
- Its panel included Judge Patrick Higginbotham, who is a senior judge of the Fifth Circuit, not the Eleventh.
- It cited other nonexistent cases, and real cases for propositions they do not contain. Zicherman, for example, is a real Supreme Court decision about the Warsaw Convention; the Eleventh Circuit version cited in "Varghese" does not exist.
- In the court's words: "Its legal analysis is gibberish."
Two other details matter for practice. Schwartz had only excerpts. When he prepared the filing he did not have the full text of any "decision," because ChatGPT produced summaries and fragments. And he asked the chatbot whether "Varghese" was real. It told him the cases were real and could be found on Westlaw and LexisNexis. A tool that invents a case will vouch for it too.
How did it happen? Schwartz testified that his firm used Fastcase, which gave limited access to federal cases, and that he turned to ChatGPT believing it was a kind of search engine. The opinion records his prompts. When asked to "provide case law," "show me specific holdings" and "show me more cases," "the chatbot complied by making them up."
What the court actually faulted
The opinion opens with the line every later court has repeated:
Technological advances are commonplace and there is nothing inherently improper about using a reliable artificial intelligence tool for assistance. But existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings.
Because the court raised sanctions on its own, the Second Circuit standard required a finding of subjective bad faith (the court cited Muhammad v. Walmart Stores East, L.P., 732 F.3d 104 (2d Cir. 2013)). Careless research alone would not have met it. As the court put it about Schwartz: "Poor and sloppy research would merely have been objectively unreasonable." The court found bad faith anyway, based on conscious avoidance and false and misleading statements, and it assessed each lawyer separately.
LoDuca, the signer. He signed the March 1 affirmation without reading a single case it cited and without asking Schwartz how the research was done, knowing Schwartz did not practice federal law. He signed the April 25 affidavit, which Schwartz drafted, with no basis for swearing to it. A cursory read would have shown that one case could not be found, that the others were excerpts, and that "Varghese" was internally inconsistent. And he told the court he was on vacation to get an extension. He was not; Schwartz was. The court called that statement knowingly false.
Schwartz, the researcher. He testified that he looked for the full "Varghese" decision and could not find it, yet said nothing in the April 25 affidavit. He offered no explanation for the missing Zicherman. His May 25 affidavit said ChatGPT "supplemented" his research, but at the hearing he agreed it was his only research. His accounts of when he asked ChatGPT whether "Varghese" was real shifted between filings. Each of those points went to bad faith.
The firm. Rule 11(c)(1) holds a firm jointly responsible for violations by its lawyers absent exceptional circumstances. The firm pointed to none, so it shared the sanction.
The court was explicit about the counterfactual. If the lawyers had come clean after Avianca's March 15 reply, or after the April orders, "the record now would look quite different." The fake citations started the problem. The silence and the cover story made it sanctionable.
Why 28 U.S.C. § 1927 did not apply
The show-cause orders had cited § 1927, which lets a court make a lawyer personally pay the excess costs of multiplying proceedings "unreasonably and vexatiously." The court declined to use it. Relying on Second Circuit precedent that the statute is aimed at dilatory tactics and delay, it found that the fake cases caused several harms, but "dilatory tactics and delay were not among them." The sanction rested on Rule 11 and, alternatively, the court's inherent power. Our Rule 11 guide explains how the three sources of sanctions authority differ.
The sanctions
The order required the respondents to:
- Mail Mata a letter enclosing the sanctions opinion, the June 8 hearing transcript and the April 25 filing with its exhibits.
- Mail the same package to each judge falsely identified as the author of a fake opinion, including the fake opinion attributed to that judge.
- File copies of those letters with the court.
- Pay a $5,000 penalty into the court's registry, jointly and severally.
The court did not order an apology, reasoning that "a compelled apology is not a sincere apology." It did not award fees, because Avianca had not asked for them. It skipped further training because the firm had already arranged mandatory CLE on AI, and it found a low likelihood of repetition.
The same day, in a separate opinion, the court granted Avianca's motion to dismiss. The Montreal Convention's two-year limit is a condition precedent to suit that is not tolled by a bankruptcy stay, as a real Eastern District of New York decision, Ireland v. AMR Corp., had already held. Mata's claim was untimely. The authority his lawyers needed did not exist because the law ran the other way.
Lessons for practice
Read what you sign. LoDuca's exposure came from signing without opening anything. If your name is on the paper, the citations are yours. A five-pass cite check is the floor.
An excerpt is not an opinion. Schwartz never had a full decision. If a tool gives you a summary or a quote, pull the opinion from a real source before you rely on it. Free sources are enough for that; see our guides to free legal research tools and CourtListener.
Do not ask the tool to check itself. ChatGPT confirmed its own inventions. Verification means comparing the citation with an independent source.
When opposing counsel says they cannot find your case, stop. That is the moment Mata turned from a research error into a sanctions case. Find the case or withdraw the cite, in writing, the same week.
Candor beats explanation. The false vacation letter and the "supplement" story did more damage than the fake cases. Later courts have drawn the same line, and our AI hallucination sanctions list shows sanctions rising when lawyers deny or deflect.
Write it down. A firm AI policy requiring that every cited authority be opened before filing, plus a 10-minute verification workflow, would have caught all of it.
Frequently asked questions
What happened in Mata v. Avianca?
A passenger sued Avianca over a knee injury on a flight to New York. His lawyers opposed the airline's motion to dismiss with a filing that cited cases ChatGPT had invented. When the airline and the court could not find them, the lawyers did not come clean. On June 22, 2023, Judge P. Kevin Castel sanctioned both lawyers and their firm, and the same day dismissed the case as time-barred.
What sanctions did the court impose in Mata v. Avianca?
A $5,000 penalty paid into the court registry, imposed jointly and severally on Peter LoDuca, Steven A. Schwartz and the Levidow, Levidow & Oberman firm. The lawyers also had to mail the client, and each judge falsely named as the author of a fake opinion, a letter attaching the sanctions opinion, the hearing transcript and the filing that contained the fake opinions. The court sanctioned under Rule 11 and, alternatively, its inherent power.
Was the lawyer sanctioned for using ChatGPT?
No. The opinion says there is nothing inherently improper about using a reliable AI tool for assistance. The court found bad faith in what followed: a signing lawyer who read none of the cited cases, a research lawyer who could not find a case and stayed silent, a false statement to obtain an extension, and misleading explanations after the court asked for the cases.
Why did the court not use 28 U.S.C. § 1927 in Mata?
The court had cited § 1927 in its show-cause orders but declined to impose a separate sanction under it. Section 1927 targets lawyers who multiply proceedings unreasonably and vexatiously, and the Second Circuit reads it as a tool against dilatory tactics. The court found that the fake cases caused several harms, but delay was not one of them.
What is the citation for the Mata v. Avianca sanctions opinion?
Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023), No. 22-cv-1461 (PKC), ECF No. 54, decided June 22, 2023. A separate opinion entered the same day, ECF No. 55, granted Avianca's motion to dismiss because the claim was filed after the Montreal Convention's two-year limit.
Check your draft before the court does
Every fake case in Mata would have failed a simple existence check before March 1, 2023. The free Hallucination Shield runs that check on your draft now: paste the text or upload the file, and it checks up to 50 citations per run against official statute and case-law sources, with no signup. It also reads each source it finds against the sentence you cited it for, which catches the real-case-wrong-proposition problem Mata had too. Run it on anything an AI tool touched.
This article is for general information and is not legal advice.