Lawyers sanctioned for ChatGPT-fabricated citations in Mata v. Avianca

Background

The matter began as an ordinary injury claim. Roberto Mata said he was hurt when a metal serving cart struck his knee during a flight operated by Avianca^[1]. He was represented by the New York firm Levidow, Levidow & Oberman, where two lawyers, Steven Schwartz and Peter LoDuca, worked on the case^[1].

What happened

While preparing the filing, Schwartz used ChatGPT to carry out legal research^[1]. The brief he produced cited judicial decisions that did not exist, among them a case named "Varghese v. China Southern Airlines"^[1]. He did not check the citations on legal databases, the usual way of confirming that a cited decision is real and actually says what it is claimed to say^[1].

The problem did not stop at the first draft. When asked whether one of the cases was real, ChatGPT answered yes, and said the cases could be found on reputable legal databases^[1]. A system that produces fluent and confident text had, in effect, vouched for its own inventions. Text-generation tools work by predicting plausible wording rather than by looking anything up, so a fabricated case name can come back looking exactly like a real one; the error is only caught if a person checks the source.

The sanction

On Thursday, June 22, 2023, Judge P. Kevin Castel of the Manhattan federal court imposed the penalty: a fine of US$5,000^[2]. Castel found that the lawyers had acted in bad faith, describing "acts of conscious avoidance and false and misleading statements to the court," and said they had abandoned their responsibilities by submitting non-existent judicial opinions and continuing to stand by them^[1].

The judge did not treat the use of AI itself as the offence. He wrote that there is "nothing inherently improper" about using a reliable artificial intelligence tool for assistance, but that existing rules impose a gatekeeping role on attorneys, who must ensure the accuracy of what they file^[1]. In other words, a tool may help draft a document, but the lawyer who signs it remains answerable for every citation in it.

The underlying case

On the same day, the judge dismissed the case itself, finding that Mata's claim had been filed after the two-year window allowed under the Montreal Convention^[2]. The claim therefore failed on its own procedural grounds, quite apart from the invented citations that had brought the lawyers into trouble.

Aftermath

The episode travelled well beyond one courtroom. Other judges, including judges in Texas, began requiring attorneys to attest that their filings were not drafted by generative AI, or that any AI-assisted drafting had been fact-checked^[1]. The case became a commonly cited example of the gap between what a language model can produce and what a court will accept as true, and of the fact that the professional duty of accuracy stays with the human who files the paper.