Every tool sold to catch AI mistakes in court filings checks whether a case is real. Nothing checks whether Officer Michelle Amarillo is real, except somebody reading the transcript, and in this appeal nobody did.
For three years the AI in court story has been about cases. A lawyer asks a chatbot for authority, the chatbot invents a decision with a plausible name and a plausible reporter cite, and a judge or opposing counsel goes looking for it and finds nothing. Whole products now exist to catch exactly that. You paste in a brief, the software checks every citation against a real database, and it flags the ones that are not there.
The New Mexico Supreme Court just dealt with the version none of those products can see.
Stephen Aarons, a private attorney in Santa Fe, was handling the appeal of Oscar Renee Sandoval. Sandoval pleaded not guilty to murdering the mother of his children, was convicted, and was sentenced last year to life in prison. The appeal is the one real chance a person in that position gets.
The brief in chief that Aarons filed described trial testimony. According to the court’s order, as reported by the Epoch Times, Aarons admitted that it contained false testimony from wholly fabricated witnesses, and the order names them: Officer Michelle Amarillo, Officer Sanchez, Manal Al-Jibury and Teresa Marquez. Reuters reported that the invented material included the kind of detail that makes a record feel solid, a shooter described as wearing dark pants and a white shirt.
Read that list again with a criminal appeal in mind. Two of the four were police officers. An appellate court reviewing a life sentence was being told what law enforcement said on the stand, and law enforcement never said it, because those officers were never there.
Aarons explained himself to the justices at a hearing on August 21, 2026. Reuters reported that he told the court he fed a computer generated transcript and other case materials to ChatGPT, expecting it to produce “a bulletproof summary.”
His account to the Epoch Times drew a line around the damage: “I wrote the brief but the table of contents and the summary contained numerous errors.” He also acknowledged he had not known that AI could hallucinate facts.
That last point is the one worth sitting with. By September 2026 the idea that chatbots invent case law is so familiar it has its own trackers, its own sanctions database and its own genre of news story. What seems to have survived in a lot of heads is the narrower belief that the invention stops at citations. That if you hand the model the real transcript, it will summarize the real transcript. It will not necessarily do that. It will produce something that reads like a summary of a transcript, and when it runs short on material it fills the gap with material that fits.
The order came down on September 9, 2026. The justices:
Aarons told Reuters: “I am remorseful but hopeful that the disciplinary board takes into account it was an honest mistake.”
On its own the $5,000 is not a large number next to what other courts have done this year, where appellate sanctions have reached $15,000 per attorney. The dollar figure is not the point of this order. Contempt and a bar on appearing in the state’s highest court are the parts that follow a lawyer around.
Every one of these stories has a client in it, and usually the client is a company that loses a motion. Here the client is a man serving a life sentence whose opening brief on appeal told the court about testimony that never happened.
The appeal is still pending. Reuters reported that it was reassigned on September 2 to public defender Kim Chavez Cook, and the Epoch Times reported that the new lawyer has been ordered to write a new brief. Whatever the merits of Sandoval’s appeal were, they now have to be argued all over again from a brief that has to be written from scratch.
This is the part that should worry anyone who thinks the problem is being handled.
A citation checker asks a yes or no question against a closed list. Does this case exist? There is a database of every reported decision, so the answer is cheap. There is no database of who testified at a particular trial in a particular county. The only source of truth is the record itself, and the only way to check a summary of testimony is to open the transcript and find the page where the witness said it.
That is slow, dull work, and it is precisely the work the tool was brought in to skip. So the safeguard for this failure is not software. It is the lawyer doing the job the summary was supposed to replace.
Four people who never existed told the New Mexico Supreme Court what happened when a woman was killed. The court caught it. The next one might be filed somewhere that does not.