ChatGPT-using lawyer punished for citing fake testimony from made-up witnesses

The New Mexico Supreme Court held attorney Stephen Aarons in direct contempt for submitting an AI-generated brief with false testimony from wholly fabricated witnesses, plus other factual and legal misrepresentations. The court referred him to a disciplinary board, fined him $5,000 payable to the State Bar Client Protection Fund, barred him from appearing before the court pending discipline, struck prior briefs, and ordered the public defender office to appoint new counsel for the defendant.

Aarons had been hired privately to appeal Oscar Renee Sandoval’s murder conviction and life sentence. In August 2025 he filed the defective brief; the state moved to strike portions weeks later. He admitted feeding a computer-generated trial transcript and other case materials into ChatGPT—he indicated a version powered by OpenAI’s o3 model—and not verifying the output before signing and filing. Fabricated or false testimony named in the order included officers and civilians who were never called, misnamed witnesses, and invented threat accounts.

This is not the familiar “fake cases” scandal alone. Aarons inaccurately described real precedents and cited invented testimony. A man in custody is the one whose appeal was delayed and whose record was polluted. That is the stake.

We’re past the era when “I didn’t know AI could hallucinate” is a credible defense in a courtroom. Justice C. Shannon Bacon said there were at least eight or nine code-of-conduct provisions violated, and she shredded the claim of ignorance—pointing to relatives who already knew the risk and to years of above-the-fold coverage of lawyers burned by AI fabrications. Chief Justice Julie Vargas and Justice Michael Vigil kept dragging the hearing back to the client in custody and to a simple rule: whatever the tool, the signer owns the brief.

Useful AI in law is real—summaries, research drafts, issue-spotting—when a human checks every cite and quote. Aarons treated a model summary as “bulletproof” because AI felt ubiquitous in legal and medical work. That is exactly how professional negligence scales: convenience plus unverified trust. The court’s contempt finding and disciplinary referral are the right response. Standing orders about AI certificates might help culture; they do not replace the duty of candor that already existed.

The harm path is clear. Fabricated testimony wastes the court’s time, hands prosecutors a strike motion, forces appointment of new counsel, and leaves a Spanish-speaking defendant—whom Aarons had not fully informed about ChatGPT’s role—waiting in custody while the mess is cleaned. Aarons’s later statement to Ars claimed remorse and an “honest mistake.” The order’s earlier finding that he demonstrated a lack of remorse and concern for his client is the institutional read that matters more than the press quote.

I’m watching whether disciplinary boards treat AI-hallucination filings as ordinary competence failures—because they are—and whether courts keep centering clients rather than turning every hearing into a policy seminar. Tools will get better. The duty to verify will not get lighter.

Ars Technica’s September 11 report by Jon Brodkin details the Wednesday contempt order, the August 21 hearing, and Aarons’s admissions about using Rev.com for transcription then ChatGPT for summarization. Many prior lawyer scandals involved fake case citations; this matter centers on fabricated witness testimony and inaccurate descriptions of real authority. Sandoval’s case is set to proceed in the court’s 2026–27 term with new appointed counsel.