He asked ChatGPT whether the cases were real. It said yes.

Roberto Mata sued Avianca over a knee injury from a metal serving cart on a flight to New York. When the airline moved to dismiss on limitation grounds, his lawyers filed a brief citing decisions that Avianca's counsel then could not find anywhere [1][2].

The brief had been researched with ChatGPT. Steven Schwartz, the attorney who wrote it, had no Westlaw or LexisNexis account; his firm used a service with limited federal coverage. He testified that he had assumed ChatGPT was "like a super search engine", and that he was "operating under the false perception that this website could not possibly be fabricating cases on its own" [1].

Six of the cited decisions did not exist. The court examined the fake "Varghese" opinion and found its legal analysis was "gibberish", that it cited other non-existent decisions, that quotation marks were left unpaired, and that it ended abruptly without a conclusion. Of Schwartz's prompts — "show me specific holdings", "show me more cases" — the court wrote simply that "the chatbot complied by making them up" [1].

Here is the part worth stopping on. Schwartz did try to verify. He asked ChatGPT directly whether the cases were real — "Is Varghese a real case" and "Are the other cases you provided fake". ChatGPT answered that Varghese "does indeed exist" and could be found on Westlaw and LexisNexis. He screenshotted that exchange and filed it as evidence of his diligence [1][2].

Peter LoDuca, the attorney of record, signed an affidavit annexing the fabricated opinions without reading them or asking a single question about them, and separately told the court he was away on vacation when he was not [1].

On 22 June 2023 Judge P. Kevin Castel sanctioned both attorneys and their firm $5,000, jointly and severally, finding subjective bad faith, and required them to write to every real judge falsely named as the author of a fabricated opinion [1][2].

Nothing here says do not use AI for legal research; the court did not say that either. What the record shows is a tool used with no ground truth attached. The same model, pointed at a database it can actually retrieve from and made to return citations that resolve, is genuinely useful — the difference is whether the answer is anchored to something a person can open. Schwartz's firm did not have that database. The absence of Westlaw is the quiet cause of this entire case [1].

The failure was not that the model invented cases; that is a known property. It is that every check Schwartz ran was inside the system that was wrong. He asked the model. The model confirmed. He asked it again after the court raised doubts, and it doubled down, naming databases it had no access to. A verification loop that never leaves the model is not verification — it is the same claim, asked twice [1][2].