A 2023 order sanctioned lawyers over ChatGPT's fake cases

What the opinion found

On 22 June 2023, Judge P. Kevin Castel of the US District Court for the Southern District of New York issued an opinion and order on sanctions in Mata v. Avianca. The order records that attorney Steven Schwartz used ChatGPT to research a brief, that the tool "fabricated the cited cases", and that Schwartz and the attorney of record, Peter LoDuca, submitted an affirmation citing six decisions - Varghese, Miller, Petersen, Shaboon, Martinez and Durden - none of which exist. The order states the fabricated Varghese opinion carried a real docket number belonging to an unrelated case and cited internal authorities that were themselves invented.

Why the court called it bad faith, not a mistake

The case docket shows the fabricated citations were first filed on 1 March 2023, and the order records that opposing counsel flagged on 15 March that the cited cases could not be located. The order finds that Schwartz and LoDuca did not withdraw the filing after that warning, after two subsequent court orders demanding the cases be produced, or after Schwartz's own attempt to verify one citation failed. The court records that LoDuca separately gave a false reason - claiming to be on vacation himself - to obtain a filing extension. Rule 11 sanctions, the order explains, require a finding of subjective bad faith when imposed by the court on its own initiative, a higher bar than ordinary carelessness.

The verification rule this implies

The order does not hold that using an AI tool for legal research is itself improper; it states plainly that "there is nothing inherently improper about using a reliable artificial intelligence tool for assistance". What it sanctions is standing behind an unverified citation after being told, more than once, that it could not be found. The court imposed a $5,000 penalty jointly on the two attorneys and their firm, and ordered letters of notice sent to the plaintiff and to each judge falsely named as an opinion's author. This is a court record of a sanction, not an allegation; the outcome and its stated reasoning are as the opinion describes them, and the same text is separately hosted in Justia's public case record.

• Can every case citation a tool produces be independently located in an actual reporter or database before it is filed anywhere?
• What is the process for withdrawing a filing once a citation's existence is challenged, and who is responsible for triggering it?
• Does a generated citation's confident tone or plausible formatting substitute for confirming the source exists?

The opinion is one court's sanction on one set of facts, not a general ruling on generative AI in litigation. Its narrower and durable point is procedural: a lawyer's duty to check a citation does not lapse because the citation came from a tool that sounded certain.