IA · 12 September 2026 · 4 min read
Lawyer relies on ChatGPT in murder appeal: court uncovers fabricated witnesses and fake police testimony
In brief: The New Mexico Supreme Court has held criminal defense lawyer Stephen Aarons in direct contempt and fined him $5,000 after he filed a murder appeal featuring witnesses and police testimony fabricated by ChatGPT. Aarons, a veteran attorney with four decades of experience, admitted he uploaded trial transcripts to OpenAI's o3 model and assumed its summary was bulletproof. The case marks an alarming escalation in courtroom AI mishaps: hallucinations are no longer confined to fake case law, but now fabricate material evidence in felony criminal trials.
by Team Mocchi's
From fabricated case law to phantom eyewitnesses
In recent years, courtrooms around the world have witnessed numerous instances of attorneys getting reprimanded for citing bogus judicial precedents concocted by large language models. However, an order issued by the New Mexico Supreme Court sets an unprecedented and far more alarming milestone: the generation not of fictitious legal doctrines, but of entirely fabricated eyewitnesses and sworn police testimony in a first-degree murder appeal.
As reported by Ars Technica, the state's highest court found defense attorney Stephen Aarons in direct contempt, slapped him with a $5,000 fine, and referred him to a disciplinary board for further professional sanctions. Aarons, who has practiced criminal defense for over 40 years, was retained to appeal the conviction of Oscar Renee Sandoval, who received a life sentence in early 2025 for the murder of his partner, Shiereen Al-Jibury.
The brief submitted by Aarons to overturn the life sentence cited testimony from people who never testified and never existed, including phantom law enforcement figures such as Officer Michelle Amarillo and Officer Sanchez, as well as alleged witnesses Manal Al-Jibury and Teresa Marquez. The filing also attributed false statements regarding the shooter's physical appearance and alleged threats to actual witnesses who had said no such thing at trial.
The OpenAI o3 breakdown and the illusion of a 'bulletproof summary'
During a disciplinary hearing, Aarons explained that he fed a digital transcript of the trial proceedings into ChatGPT alongside other case files, operating under the assumption that the tool would provide an accurate, ready-to-file recap of the record. As detailed by The Verge, Aarons admitted relying on a build powered by OpenAI's o3 model, believing it would generate a «bulletproof summary» of the murder trial.
The justices offered an excoriating assessment of his conduct. Justice C. Shannon Bacon challenged the attorney directly: «Counsel, do you watch the news? Do you listen to the radio? Do you read anything about what’s going on in the world? Because the problem with lawyers relying on AI hallucinations is an above-the-fold story every single day.» The court noted that Aarons failed to perform any cross-checking against the original record before filing the brief, and never informed his incarcerated client that the submission was riddled with misrepresentations.
When hallucination shifts from legal theory to factual fabrications
Previous high-profile clashes between generative AI and the judiciary—such as the infamous New York case Mata v. Avianca—primarily involved lawyers citing phantom case citations to bolster legal briefs. The New Mexico incident crosses a critical line: the model did not merely misinterpret precedents, but hallucinated the primary factual bedrock of a criminal trial, creating fictional witnesses and false police statements when a man's life in prison hung in the balance.
This breakdown illustrates a common conceptual trap in casual enterprise adoption of frontier LLMs. Despite expanded context windows and enhanced reasoning capabilities, large language models remain probabilistic text generators. When tasked with unstructured synthesis without strict retrieval constraints, they frequently fill informational gaps with syntactically convincing yet entirely fictional data. In high-stakes environments, relying on unchecked model generation constitutes gross negligence.
Mocchi's take
This case serves as a stark reminder of the perils inherent in treating frontier generative models as plug-and-play document synthesizers in regulated industries. For organizations and enterprises looking to leverage AI across complex technical documentation, procurement files, or compliance contracts, Aarons' failure offers an essential operational lesson. Critical workflows cannot rely on raw conversational prompts: they require rigorously engineered Retrieval-Augmented Generation (RAG) architectures with deterministic citation validation and an unyielding human-in-the-loop verification protocol. Delegating factual truth to an unconstrained model without strict safeguards is not an efficiency gain; it is an open operational liability.