The New Mexico Supreme Court sanctioned a lawyer for submitting an appeal brief that included fabricated testimony and witnesses created by ChatGPT. The court took issue not with the use of AI itself, but with the lawyer signing and filing the document without verifying the output.

Ars Technica reported on Sept. 11, local time, that the New Mexico Supreme Court held criminal defence lawyer Stephen Aarons (스티븐 애런스) in contempt and ordered him to pay $5,000 to the New Mexico State Bar client protection fund. Aarons was also referred to a disciplinary committee and was barred from appearing before the state Supreme Court until the related investigation and procedures are completed.

The case involved an appeal by Oscar Renee Sandoval (오스카 르네 산도발), who was sentenced to life in prison for murder. In a filing submitted in August last year, Aarons listed four witnesses who did not appear at the actual trial: officer Michelle Amarillo (미셸 아마릴로), officer Sanchez (산체스), Manal Al-Jibury (마날 알지버리) and Teresa Marquez (테리사 마르케스). He also added threats that did not exist to testimony by actual witnesses including Danny (대니) and Linda Stanton (린다 스탠턴), and included false statements about the suspect's clothing and appearance. He also described parts of some existing precedents inaccurately.

At a hearing on Aug. 21, Aarons testified that he had entered trial transcripts and case materials into ChatGPT to have them summarised and believed the results would be accurate. He argued that he did not sufficiently understand the hallucination problem in which AI fabricates facts.

The panel did not accept that explanation. It said it did not matter whether the help came from AI, a junior lawyer or a colleague, and that the lawyer who signs a filing bears the final responsibility to confirm its accuracy. The court also pointed to Aarons not only failing to identify errors but also not properly informing his client of the seriousness of the issue. The Supreme Court struck all of the existing appeal filings from the record and ordered the public defender's office to appoint a new lawyer. Sandoval's appeal is to be heard again in the 2026 to 2027 term after a new filing is submitted.

The decision is viewed as reaffirming that generative AI can be used in legal work, but responsibility for verifying the output cannot be shifted to AI.