A New Mexico defense attorney has been fined $5,000 and held in contempt after submitting a murder appeal containing something courts have increasingly warned lawyers about: information invented by artificial intelligence.

But this case went beyond the fake legal citations seen in several previous AI-related court controversies. The brief contained testimony attributed to witnesses who did not exist, along with inaccurate statements supposedly made by real witnesses and police officers. Attorney Stephen Aarons acknowledged using ChatGPT to help prepare the filing and admitted he did not verify the AI-generated material before submitting it.

What Happened?

Stephen D. Aarons, a private criminal defense attorney in Santa Fe, was representing Oscar Renee Sandoval in an appeal before the New Mexico Supreme Court.

Sandoval had been convicted of murdering Shiereen Al-Jibury, his partner and the mother of his children, and was sentenced to life in prison. Aarons was hired to handle the appeal.

While preparing the case, Aarons used ChatGPT to process a computer-generated transcript of the murder trial along with other case materials.

He expected the AI system to help create what he described during a court hearing as a “bulletproof summary” of the proceedings.

Instead, the resulting brief contained information that was never part of the trial.

ChatGPT Invented Entire Witnesses

The most serious problem was not simply an incorrectly summarized court decision.

According to the New Mexico Supreme Court’s order, the brief included false testimony attributed to four entirely fabricated witnesses: Officer Michelle Amarillo, Officer Sanchez, Manal Al-Jibury and Teresa Marquez.

Those people had not testified at the trial.

The brief also included false or inaccurate testimony attributed to people who did exist.

Among the errors were claims about threats allegedly received by witnesses and statements concerning the shooter’s clothing and appearance.

The filing additionally misrepresented legal authority from previous New Mexico cases.

One Fabricated Detail Described the Shooter’s Clothes

The Supreme Court had already ordered Aarons to explain himself after discovering suspicious information in the appeal.

One of the invented details involved statements describing the shooter as wearing dark pants and a white shirt.

The problem was that the testimony supporting that description did not exist in the trial record in the form presented by the brief.

The court ultimately concluded that the filing contained both fabricated factual material and inaccurate legal representations.

Aarons Admitted He Didn’t Verify the AI’s Work

Perhaps the most important part of the case was what happened after ChatGPT produced the material.

Aarons admitted that he did not independently verify the factual claims and legal authorities before signing and submitting the brief.

He also acknowledged that he did not inform his client that the filing contained these problems.

That failure turned an AI error into a professional responsibility issue.

Generative AI systems can produce fluent, confident answers even when some of the information is inaccurate or completely fabricated. This behavior is commonly called an AI “hallucination.”

Aarons said he had not understood the extent to which that could happen.

A Supreme Court Justice Was Stunned by the Explanation

During an August 21 hearing, the New Mexico Supreme Court questioned Aarons about how the errors had reached the court.

Justice C. Shannon Bacon challenged his claim that he did not fully understand the risks associated with AI-generated legal material, pointing out how widely AI hallucinations have already been discussed publicly.

Aarons acknowledged his mistake during the hearing.

He explained that because AI tools were becoming widely used in professional fields, he had assumed the output generated from the trial materials would provide a reliable summary.

That assumption proved costly.

The Court Fined Him $5,000

The New Mexico Supreme Court held Aarons in direct contempt.

It ordered him to pay a $5,000 fine to the State Bar of New Mexico Client Protection Fund.

The court also said Aarons would be referred to the attorney disciplinary board for further investigation. Additional professional consequences could therefore follow.

The justices were particularly critical of his handling of the situation, concluding that he had shown insufficient remorse and concern for his client.

Aarons later said he was remorseful and hoped disciplinary authorities would recognize that what happened was an honest mistake.

He Has Also Been Barred From Appearing Before the Court

The financial penalty is not the only consequence.

Aarons has been barred from appearing before the New Mexico Supreme Court while disciplinary proceedings remain unresolved.

The court also struck the previous briefs in Sandoval’s appeal from the record.

A new attorney has been assigned to represent Sandoval, allowing his appeal to continue without relying on the compromised filings.

Public defender Kim Chavez Cook was assigned to the appeal on September 2.

The underlying murder appeal therefore remains pending.

This AI Mistake Was More Serious Than Fake Case Citations

Lawyers getting into trouble for AI hallucinations is no longer unprecedented.

Courts in the United States have sanctioned attorneys after AI-generated legal briefs included cases that did not exist, incorrect quotations or inaccurate descriptions of real decisions.

What makes the New Mexico case particularly striking is the nature of the hallucination.

ChatGPT did not merely produce a nonexistent court citation.

The filing contained fictional witnesses and testimony presented as if they were part of a real murder trial.

In a criminal appeal involving a defendant serving a life sentence, inaccurate representations of trial evidence can have especially serious consequences.

Why AI Can Produce Convincing False Information

Large language models such as ChatGPT generate responses by predicting and constructing language based on patterns learned during training and information available in the conversation or supplied context.

They are capable of summarizing documents, organizing information, explaining concepts and assisting with drafts.

But fluent language does not automatically guarantee factual accuracy.

A model can sometimes generate information that sounds plausible even when that information is unsupported.

That makes verification particularly important in fields such as law, medicine, finance and academic research, where a seemingly small factual error can materially change an outcome.

The New Mexico case demonstrates the problem clearly: fabricated testimony can still sound like legitimate courtroom testimony when embedded naturally inside a polished legal document.

AI Wasn’t the Only Failure

The case also highlights an important distinction in discussions about AI mistakes.

ChatGPT generated inaccurate information, but the attorney was responsible for the document ultimately submitted to the court.

The New Mexico Supreme Court focused heavily on Aarons’ failure to verify the material before signing and filing it.

That principle matters because lawyers have professional responsibilities that cannot simply be delegated to software.

AI can assist with research, summarization and drafting, but the person submitting a legal document remains responsible for checking whether its claims, quotations, citations and evidence are accurate.

A Warning for Professionals Using Generative AI

The controversy is another reminder that generative AI can be extremely useful without being an authoritative source of truth.

For lawyers, developers, researchers, journalists and other professionals, the safest approach is to treat AI-generated material as a draft that requires verification rather than finished evidence.

That means checking important claims against primary sources, opening cited cases rather than assuming they exist, comparing summaries with original documents and confirming quotations before publishing or submitting them.

Aarons believed ChatGPT would give him a reliable summary of a murder trial.

Instead, fictional witnesses entered a real Supreme Court filing.

The result was a $5,000 fine, a contempt finding, a disciplinary referral, removal from his client’s appeal and another high-profile example of what can happen when convincing AI output is mistaken for verified fact.