The New Mexico Supreme Court has fined defense attorney Stephen Aarons $5,000, held him in contempt and said it would refer him for disciplinary investigation after an AI-assisted appellate brief in a murder case included invented witnesses and false testimony. The sanction, reported September 11, turns a familiar generative-AI failure mode—plausible but false output—into a high-stakes example of professional accountability in a criminal appeal.
The court said Aarons’ filing “contained false testimony from wholly fabricated witnesses,” according to reporting on the order. The brief concerned the pending appeal of Oscar Renee Sandoval’s murder conviction. Aarons did not verify the factual claims and legal authority in the document before signing and filing it, reports said.
How the filing failed
Aarons told Reuters that he had used ChatGPT to summarize trial proceedings after taking on the appeal and that he did not understand the extent to which the system could hallucinate facts. At an August 21 hearing, he said he had expected the tool to create a “bulletproof summary” after he provided a computer-generated transcript and other case materials, Reuters reported. The court’s action followed that hearing; reporting describes the contempt order as dated September 9.
The distinction between drafting help and filed work is central to the court’s response. The reported issue was not simply that a lawyer used an AI system to organize material or prepare an initial draft. It was that factual assertions and authorities generated or incorporated through that process were submitted to a court without being checked against the record and the law. For a criminal defendant, errors of that kind can affect the handling of an appeal; for a lawyer, they can trigger sanctions and a disciplinary review.
The appeal itself remains pending. On September 2, it was assigned to New Mexico public defender Kim Chavez Cook, Reuters reported. The reassignment separates the continuing appellate matter from the court’s action against Aarons, but it also underscores the practical cost of a defective filing: time and legal work may be required to restore a reliable presentation of the case.
Verification duties in AI-assisted legal work
The ruling adds a more severe factual dimension to prior disputes over AI-generated court submissions. Many earlier cases centered on nonexistent judicial decisions, inaccurate quotations or authorities cited for propositions they did not support. Here, the reported defects included purported witness testimony that did not exist. That makes record verification—not only citation checking—the central governance lesson for legal teams that use generative AI in research, summarization and drafting.
That lesson is consistent with recent American Bar Association analysis of AI-related litigation sanctions. An ABA Litigation Section article says that a lawyer whose name appears on a filing remains responsible for the accuracy of legal authorities, whether an error comes from delegation, carelessness or reliance on generative AI. The article also describes sanctions that can extend beyond fines, including referrals to disciplinary authorities and other court-imposed remedies. (americanbar.org)
For organizations deploying generative AI in regulated work, the New Mexico case illustrates why a general instruction to “review AI output” is not enough. Controls need to match the consequence of the document: a human reviewer must compare assertions to the source record, confirm that cited authorities exist and support the stated point, and ensure that a signer can stand behind the final submission. An AI system can accelerate the creation of a summary, but it cannot assume the professional duty attached to a lawyer’s signature.
The case also does not establish that generative AI has no legitimate role in legal practice. ABA materials describe potential uses for AI tools in tasks such as organizing documents, summarizing pleadings and assisting with research, while emphasizing that users must account for errors and independently verify output. (americanbar.org) The New Mexico sanction instead draws a bright operational line: assistance from a generative model does not excuse unverified material in a court filing.
Accountability at the point of filing
As courts, law firms and legal-technology providers build policies around AI-assisted work, the ruling offers a concrete standard for the highest-risk stage of the workflow. The final filer remains accountable. In this case, that accountability resulted in contempt, a $5,000 fine and a disciplinary referral—not because an AI tool was used, but because material attributed to that process was filed without the verification required of counsel.
