Steven Schwartz's ChatGPT brief cited six court cases that didn't exist

The brief that cited six cases no one could find

Roberto Mata sued Avianca, Inc. after a metal serving cart struck his knee on a 2019 flight. He filed his complaint in a New York state court on or about February 2, 2022; Avianca removed the case to the U.S. District Court for the Southern District of New York three weeks later. In January 2023 it moved to dismiss, arguing Mata had waited too long to sue under the two-year deadline the Montreal Convention sets for international-flight injury claims.¹²

Opposing that motion fell to Steven A. Schwartz, an associate at the plaintiff's firm, Levidow, Levidow & Oberman, who had practiced law for roughly three decades.¹ Schwartz used ChatGPT to research the opposition brief. He asked it to argue the statute of limitations was tolled by the defendant's bankruptcy, then to provide supporting case law. When the results felt thin, he asked for more: to "show me specific holdings in federal cases," to "show me more cases," to "give me some cases."¹ ChatGPT complied every time. The firm had no independent way to check what it returned: Levidow, Levidow & Oberman did not subscribe to Westlaw or LexisNexis, only to Fastcase.¹

Peter LoDuca, the attorney of record, signed and filed the resulting brief on March 1, 2023. It cited six decisions that do not exist: 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. Each carried a docket number and a reporter citation. Three of the six, presented as federal appellate and district rulings, quoted language attributed to real federal judges.¹ Avianca's lawyers went looking for the cases and could not find seven of the decisions cited in the brief. Their March 15 reply said so, and the court, unable to locate the same authorities on its own search, ordered Schwartz and LoDuca to produce copies.¹ Judge P. Kevin Castel ordered them produced on April 11 and again on April 12.¹

Told to prove it, the chatbot vouched for itself

Ordered to produce the cases, Schwartz went back to the tool that had supplied them. On April 25 he filed an affidavit annexing what he said were copies of the decisions, including a fake Eleventh Circuit opinion in Varghese attributed to Judges Adalberto Jordan, Robin Rosenbaum, and Patrick Higginbotham.¹ The fake opinion reads like a real one until it doesn't. It opens with a wrongful-death claim brought by Susan Varghese for her deceased husband. Within the same paragraph it becomes an overbooking dispute brought by a different Varghese, Anish, over a missed connection in Guangzhou. Judge Castel's order found "The 'Varghese' decision shows stylistic and reasoning flaws that do not generally appear in decisions issued by United States Courts of Appeals," and put it more bluntly two sentences later: "Its legal analysis is gibberish."¹

Before he filed the affidavit, Schwartz had asked ChatGPT directly: "Is varghese a real case." It answered: "Yes, Varghese v. China Southern Airlines Co Ltd, 925 F.3d 1339 (11th Cir. 2019) is a real case." He asked for a source. It apologized "for the confusion earlier" and insisted the case "does indeed exist and can be found on legal research databases such as Westlaw and LexisNexis." He asked whether the other cases he'd been given were fake too. It told him no, that they were "real and can be found in reputable legal databases."¹

Schwartz's affidavit called the fake decisions genuine, and it did not disclose that a chatbot had generated them. The truth came out only after Judge Castel ordered the lawyers to show cause, on May 4, why they should not be sanctioned. Schwartz and LoDuca began, in the court's words, to "dribble out the truth" starting May 25. That day, Schwartz submitted the screenshots above and acknowledged that six of the decisions — Varghese, Miller, Petersen, Shaboon, Martinez, and Durden — were generated by ChatGPT and do not exist.¹ Of the seven decisions Avianca's lawyers had flagged as unlocatable in March, one turned out to be a real precedent, cited alongside the fake ones; the other six were pure invention. The New York Times reported the disclosure on May 27, and the case Schwartz had hoped to win on a procedural technicality became, instead, a national story about lawyers trusting a chatbot.³

The sanction was never about using ChatGPT

A large language model like ChatGPT does not look anything up. It generates the next word of an answer from patterns in its training text, aiming for language that reads as a plausible continuation of the prompt in front of it. That is why a model builds a fake citation the same way it builds a true one. Asked for case law supporting a legal argument, it produced text shaped exactly like a citation: a party name, a reporter volume, a court, a year, a quoted holding. Nothing in that process checks the case against a legal database, because the model has no database to check against. Told the case was fake and asked again, it produced text shaped like a confident correction rather than a retraction, because a confident answer was still the most plausible thing to say next. That is why "Is varghese a real case" got "Yes" instead of "I don't know": the model has no setting for either.

That explains why ChatGPT produced six invented cases. Nobody at Levidow, Levidow & Oberman opened a case reporter, or a free legal search, to check them before filing. Schwartz testified he was "operating under the false perception that this website could not possibly be fabricating cases on its own" and that he "just never thought it could be made up."¹ He described the tool, in his own words, as "a super search engine," and misread the reporter abbreviation "F.3d" as standing for "federal district, third department."¹ What Judge Castel's order sanctions is what came after that misunderstanding: filing an affidavit with no basis for believing it was true, and then giving the court, in the order's words, "shifting and contradictory explanations."¹ Schwartz's own affidavit described ChatGPT as something that had "supplemented" his research; under questioning at the June 8 hearing, he agreed it had not supplemented anything. "THE COURT: But ChatGPT was not supplementing your research. It was your research, correct? MR. SCHWARTZ: Correct. It became my last resort."¹ LoDuca, who signed and filed the brief, separately told the court he needed a scheduling extension because he was on vacation. He was not, and knew the statement was false when he made it.¹

On those facts, Judge Castel found bad faith "based upon acts of conscious avoidance and false and misleading statements to the Court," not on the choice to consult an AI tool.¹ On June 22, 2023, he sanctioned Schwartz, LoDuca, and the Levidow firm $5,000, paid jointly into the court's registry within 14 days. He also ordered each of them to send a letter to Roberto Mata and to every real judge falsely credited with authoring a fake opinion.¹ He declined to pursue forgery charges, since the fake opinions carried no signature or seal, and stopped short of a heavier penalty, noting the fabrications were not made for financial gain and that neither lawyer had a disciplinary history.¹ Reuters and other wire services covered the ruling as a cautionary tale about lawyers trusting a chatbot.⁴

Two more courts, two more fabricated briefs

The Mata sanction did not stop the pattern. Seven months later, in Park v. Kim, the Second Circuit found that an attorney's reply brief, drafted with ChatGPT's help, cited a case, Bourguignon v. Coordinated Behavioral Health Services, that does not exist. The panel wrote that "citation in a brief to a non-existent case suggests conduct that falls below the basic obligations of counsel" and referred the attorney to its grievance panel.⁵ In February 2025, a federal judge in Wyoming, Kelly Rankin, sanctioned three lawyers at Morgan & Morgan after a motion cited nine cases, eight of them invented by MX2.law, the firm's own in-house AI research tool. The penalties again totaled $5,000, split between the drafting attorney, his supervising partner, and local counsel; the drafter also lost his standing to appear before that court.⁶ Judge Rankin's order called verifying legal authority "a nondelegable responsibility," whether the source is another attorney or an AI tool. A law professor who covered the ruling quoted her adding a plainer version of the same point: "The researcher must still read the case to ensure the excerpt is existing law to support their propositions and arguments."⁶⁷

Both cases are entries in a database a French law professor, Damien Charlotin, maintains of court rulings on AI-hallucinated filings worldwide. By the middle of 2026 it listed well over a thousand documented cases, a number that keeps climbing.⁸ Six invented cases in one federal brief was national news in 2023. By 2026 it is a category with its own tracker.