OddsFrontNew Mexico Supreme Court sanctions lawyer over AI-fabricated witnesses
The contempt finding moves the legal profession’s generative-AI problem beyond false citations, showing how unchecked model output can contaminate the factual record of a criminal appeal.
New Mexico’s Supreme Court has fined attorney Stephen Aarons $5,000 and held him in contempt after a criminal-appeal brief contained police testimony and witnesses that did not exist. The court said the false material appeared in work prepared with ChatGPT and ordered that the matter be referred to the state’s attorney disciplinary system. The underlying murder appeal remains pending and has been reassigned to a public defender.
A factual failure, not merely a bad citation
Previous disciplinary cases involving generative AI have often centred on invented precedents, inaccurate quotations or nonexistent case numbers. This episode is more serious because the model-generated material was presented as part of the evidentiary history of a murder prosecution. According to the court’s account, the brief included fictional witness statements and details about the alleged shooter’s clothing, potentially confusing the record on which an appellate court must evaluate a conviction.
Aarons told the court that he had supplied a computer-generated trial transcript and other materials to ChatGPT and expected it to produce a dependable summary. That workflow failed at the most basic verification stage: generated prose was treated as a faithful account of source documents without being checked line by line. The justices concluded that the conduct justified both a financial penalty and a formal contempt finding.
Why the ruling matters beyond one case
The sanction illustrates the difference between using AI as an administrative aid and delegating professional judgment to it. Language models generate plausible sequences of words; they do not independently certify that a witness, quotation or event appears in a court record. In criminal proceedings, where a defendant’s liberty and the integrity of a conviction are at stake, fabricated facts can undermine due process even if they are introduced by mistake.
The United Nations has separately identified automation bias, transparency and accountability as central risks when AI enters justice systems. Its 2026 discussion on AI and justice recommended practical safeguards such as impact assessments, audit mechanisms and clear human oversight. The New Mexico case supplies a concrete example of why those protections cannot be limited to software formally purchased by courts: lawyers’ private use of consumer AI tools can also affect judicial decision-making.
The disciplinary board will now determine whether further professional penalties are warranted. The pending appeal must meanwhile be assessed from an authenticated record, without relying on the fabricated material. Courts and bar authorities elsewhere will be watching whether New Mexico’s response becomes a model for mandatory disclosure, verification duties or training requirements when lawyers use generative AI.