Schwartz was using ChatGPT for legal research for the first time when he put it to work drafting the ten-page brief he hoped would convince Manhattan Federal Judge P. Kevin Castel not to dismiss his case, he told the court in an affidavit on Thursday, explaining that he “therefore was unaware of the possibility that its content could be false.”
When asked, ChatGPT even told Schwartz – a lawyer with 30 years of experience – that the half dozen cases it cited in the legal submission were real, he insisted. Declaring he “greatly regrets” putting his faith in the large language model, he promised to “never do so in the future” – at least, not “without absolute verification of its authenticity.”
Schwartz’s law firm Levidow, Levidow & Oberman was representing airline passenger Roberto Mata in a personal injury lawsuit against airline Avianca regarding an incident on a 2019 flight. When the airline responded to the suit by filing for dismissal, arguing the statute of limitations had expired, Schwartz and his firm responded with the ChatGPT-addled brief. Avianca’s lawyers complained to the judge that the cases cited didn’t exist, but when Judge Castel ordered Mata’s lawyers to provide copies of the questionable opinions, they promptly did so – only for Avianca’s attorneys to retort that no such cases appeared in real-life court dockets or legal databases.
Judge Castel responded earlier this month with an order demanding Schwartz and his colleagues show cause as to why they should not face disciplinary sanctions for using a chatbot to write his legal brief. The hearing is scheduled for June 8.
In response, Schwartz insisted in an affidavit filed on Thursday that he had performed all of the legal research found in the questionable brief, merely deploying ChatGPT to “supplement” his own work. “In consultation” with the AI software model, he came upon the bogus cases, and “ChatGPT assured the reliability of its content,” he explained. He even attached a transcript of his conversation with the chatbot, which apparently outwitted him by answering questions such as “are the other cases you provided fake” with “no, the other cases I provided are real and can be found in reputable legal databases.”
While ChatGPT’s responses to user queries often appear factual, the large language model functions as a probability engine, populating a text string based on the contents of its massive database of text snippets.