The New Mexico Supreme Court has sanctioned a defence lawyer whose appellate brief contained, in the court's words, "false testimony of entirely fabricated witnesses." The lawyer, Stephen Aarons of Santa Fe, says he used ChatGPT to summarise the trial record in an appeal he took on last year, and that he did not understand how readily the technology could hallucinate facts. The justices added that he had shown a lack of remorse and a lack of concern for his client — who is serving a life sentence for murder, and whose appeal is still pending.
Dozens of American lawyers have now been disciplined for filings in which a model invented case citations or misquoted statutes. Aarons's case is a different failure, and a worse one. A fabricated citation is checkable in seconds against a database that either contains the case or does not. Fabricated testimony is checkable only against the trial record — the exact document Aarons had handed to the model to read for him.
The detail the court singled out shows how ordinary the invention looked. Last month the justices ordered Aarons to explain how the material reached the principal appellate brief, noting it included "fictitious claims that the shooter wore dark trousers and a white shirt." That is not a legal error dressed up as an authority. It is evidence, of the kind that decides who was identified and who was not, produced out of nothing and filed in a murder appeal.
At an August 21 hearing, Aarons said he had uploaded a computer-generated transcript and other case materials into ChatGPT and expected an "impeccable summary." The justices were openly sceptical that a practising attorney could be unaware the tool makes things up. Justice Shannon Bacon asked whether he watches the news, listens to the radio, or reads anything about what is happening in the world, saying the problem of lawyers relying on AI hallucinations is on the front pages every day.
In a statement, Aarons said he regretted what happened and hoped the disciplinary board would treat it as an honest mistake, calling it a lesson for every professional who relies on powerful but sometimes unstable technology. OpenAI did not immediately respond to a request for comment.
The plea is weaker than it sounds, and Bacon's question is why. "I did not know it hallucinated" was a defensible position for a short window; it stopped being one while this appeal was being drafted. The more interesting question is not the lawyer's ignorance but his workflow. He fed a machine-produced transcript to a machine and asked for a summary — two layers of automation over a record nobody had read end to end, with no step at which a human compared the output to the source. No rule about AI use would have caught that. A rule about reading your own record would.
It is also worth asking why OpenAI is the party being asked for comment. Treating the vendor as a participant in a court sanction flatters the idea that the model acted. The sharper question is a product one: a system marketed as a summariser was given a transcript and returned assertions that were not in it, and the industry has no accepted name for that failure beyond a euphemism borrowed from psychiatry.
The costs here are already distributed, and not evenly. Aarons faces a disciplinary board. Oscar Rene Sandoval — convicted of killing the mother of his children, who did not plead guilty and received a life sentence last year — had his one appeal filed with invented facts in it. On September 2 the case was transferred to New Mexico public defender Kim Chavez Cook, who declined to comment, as did local prosecutors. Whatever the board decides about the honest mistake, someone now has to rebuild that appeal from the record, and the man waiting on it has lost a year he cannot get back.