A Murder Appeal Brief Contained Witnesses That Never Existed, and the Lawyer Paid for It
What happened
Stephen Aarons, an attorney in New Mexico, submitted a murder appeal brief to the state Supreme Court that contained police testimony and witnesses that did not exist. The court found the fabrications, and Aarons was fined and held in contempt.
The brief had been prepared with the assistance of OpenAI's ChatGPT. The model, prompted to help construct arguments for the appeal, did what generative language systems routinely do with legal research: it produced plausible-sounding citations, witness accounts, and factual claims that had no basis in any actual case record. Aarons filed the document without verifying whether the people named in it or the statements attributed to them corresponded to anything real.
The New Mexico Supreme Court responded with a $5,000 fine and a contempt citation. The contempt turned on a specific failure, not that the model had invented content, but that Aarons had passed that invented content to the court as if it were verified fact. Generative text systems produce what fits the context, with no inherent obligation to accuracy, and no existing rule in most jurisdictions requires a practitioner to disclose when one was used or to certify that its outputs were checked against primary sources before filing.
This is not an isolated case. Federal courts have sanctioned attorneys in other jurisdictions for the same pattern of failure since at least 2023: a practitioner uses a language model to accelerate research or drafting, the model fabricates names, case citations, or testimony, and nobody confirms the output before the document is filed. The New Mexico matter carries particular weight because it arose in a murder appeal, where a fabricated record does not merely embarrass the filing attorney. It introduces false facts into proceedings that determine whether a person remains incarcerated.
What the Aarons case makes visible is a structural gap between how legal documents are produced and how they are verified. Courts operate on the assumption that what an attorney files reflects actual evidence because an officer of the court is supposed to have confirmed it. When a language model inserts invented witnesses into a brief, that assumption breaks at the point of drafting, before the court has any chance to catch it. A provable record of what a system produced, who reviewed it against source material, and when that review occurred, would close that gap at the right moment. Without it, the obligation to catch AI-fabricated content falls on opposing counsel or on the court itself, both of whom encounter the error after the document has already been submitted.
Reported impact
- Affected parties
- Not publicly disclosed
- Harm type
- Not publicly disclosed
- Scale
- Not publicly disclosed
- Financial impact
- Not publicly disclosed
- Regulatory action
- Not publicly disclosed
Classification
Relevant governance controls
Governance control mapping is not available for this record.
- No controls mapped
Not publicly disclosed
Control mapping is analytical. It does not state that any control would have prevented the incident.
Sources and evidence
This record was researched and written by the Index. The event is also catalogued in the following database, which is listed for cross-reference.