Mata v. Avianca: what happened when lawyers cited fake ChatGPT cases?

Quick answer In Mata v. Avianca, a personal injury case against the airline in federal court in Manhattan, lawyers filed a brief in March 2023 citing court decisions that ChatGPT had invented. On June 22, 2023, Judge P. Kevin Castel sanctioned attorneys Steven A. Schwartz and Peter LoDuca and their firm, imposing a $5,000 penalty and finding bad faith.

What happened in Mata v. Avianca?

The case itself is ordinary. Roberto Mata says a metal serving cart struck his left knee on a flight from El Salvador to New York's JFK airport. He sues Avianca. The airline moves to dismiss, arguing the claim is time-barred under the Montreal Convention, the treaty that governs international air travel claims.¹

Mata's lawyer, Steven A. Schwartz of Levidow, Levidow & Oberman, needs cases showing the deadline was paused by the airline's bankruptcy. His firm's research service has limited access to federal cases. So he tries something new.¹

He later told the court he had heard about a new site that he "falsely assumed was like a super search engine called ChatGPT, and that's what I used."¹ He had never used ChatGPT before.¹

He asks it to argue that bankruptcy tolls the limitations period. Then for case law. Then: "show me more cases."¹ ChatGPT obliges. It produces cases with names, citations, judges and quotes. On March 1, 2023, a filing signed by his colleague Peter LoDuca, the attorney of record, cites them to the court.¹

Two weeks later, Avianca's lawyers reply. They say they have been "unable to locate most of the case law cited," and they list seven purported decisions by name.¹

In April, the court orders the lawyers to produce the cases. Screenshots Schwartz later filed show he had asked ChatGPT "Is Varghese a real case" and "Are the other cases you provided fake." ChatGPT answered that it had supplied "real" authorities that could be found through Westlaw, LexisNexis and the Federal Reporter.¹

They were not real. The lawyers acknowledged that six of the decisions, including "Varghese," were generated by ChatGPT and do not exist.¹

Its legal analysis is gibberish.

Judge P. Kevin Castel on the fake "Varghese" decision, Opinion and Order on Sanctions, June 22, 2023

What did Judge Castel decide?

The judge was clear that the tool was not the core problem. "Technological advances are commonplace and there is nothing inherently improper about using a reliable artificial intelligence tool for assistance," he wrote. "But existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings."¹

What turned a mistake into sanctions was what came after. The lawyers, the opinion says, "abandoned their responsibilities when they submitted non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT, then continued to stand by the fake opinions after judicial orders called their existence into question."¹ If they had come clean after Avianca's March 15 brief, or after the court's April orders, "the record now would look quite different."¹ The first admission did not come until May 25.¹

The court found bad faith on the part of both individual lawyers, based on "acts of conscious avoidance and false and misleading statements to the Court."¹ LoDuca had signed the filing without reviewing any of the cases cited in it.¹

The sanctions: a $5,000 penalty imposed jointly and severally on Schwartz, LoDuca and the firm, plus letters to their client and to each judge falsely named as the author of a fake opinion, each with a copy of the sanctions order attached.¹ The court did not require an apology, noting that "a compelled apology is not a sincere apology."¹ The firm said in a statement that it "respectfully" disagreed that its lawyers had acted in bad faith: "We made a good-faith mistake in failing to believe that a piece of technology could be making up cases out of whole cloth."² In a separate opinion, the judge threw out Mata's underlying claim because the statute of limitations had expired.²

Why it happened

ChatGPT is a language model. It produces text that sounds right, and a court citation is just a pattern of text. The Guardian explained that chatbots like ChatGPT can be prone to "hallucinations," and that their human-seeming responses can convince users an answer is correct.² Asking the chatbot whether its own cases were real did not help. It said yes.

Schwartz himself described the trap in a court filing: he "could not fathom that ChatGPT could produce multiple fictitious cases, all of which had various indicia of reliability such as case captions, the names of the judges from the correct locations, and detailed fact patterns and legal analysis that sounded authentic."¹

But the model was only one link in the chain. A second lawyer signed without reading the cases. Schwartz testified that when he looked for "Varghese" he "couldn't find it," yet the April 25 affidavit did not say so. And when the other side raised the alarm, the lawyers did not stop and check. The judge wrote that reading only the opening passages of the fake "Varghese" decision would have revealed that it was "internally inconsistent and nonsensical."¹

What this means for your business

You may never file a legal brief. But you probably publish product claims, policy pages, supplier emails and blog posts, and more of that text now starts in an AI tool. The Mata lessons carry over:

• Check facts against a real source, not the AI. If a draft cites a study, a law, a price or a statistic, open the original. Asking the chatbot "is this real?" is not a check.
• The person who signs is responsible. LoDuca signed without reading the cases. In your business, whoever hits publish or send owns the content.
• When someone flags an error, stop and verify. The worst part of this case was doubling down. A quick correction is cheap; defending a mistake is expensive.
• Keep a record. Schwartz's saved ChatGPT history showed the court exactly what he asked. Logs of who did what help you find and fix problems fast.

Air Canada learned a similar lesson when its chatbot invented a refund rule: see the Air Canada chatbot ruling. For the human check that catches these problems, read why the confirm button matters.

How ZNOS handles this

The way we run it, an AI employee drafts and a person decides what goes out. In ZNOS, each AI employee owns a job, such as Tanmay on growth and Suri on design, and works from the company's live data through the Brain rather than from guesses.

Content drafted by an agent waits for human approval before publishing. More broadly, autonomy is earned gradually: routine decisions are handled, and risky ones (money, customers, anything published under your name, production changes) wait for a human approval. An Audit Log records who changed what, so you can trace any problem back to where it started. You can meet the ZNOS team, or read more real AI incidents.