Fictional Witnesses in a Murder Appeal

A defense attorney handling a murder-conviction appeal fed a case transcript and related materials to ChatGPT and asked for a summary to help draft his brief. The summary that came back described police testimony and eyewitness statements that were never given, including a detailed, invented account of what the shooter was supposedly wearing. The attorney did not check the summary against the actual trial record before filing it with the New Mexico Supreme Court. The fabricated material was only caught when the court itself reviewed the brief.

A garden-variety summarization error, considered, since summarizing a long transcript is a hard task and some loss of detail is normal. Ruled out because this wasn't compression or omission; it was the invention of specific testimony and witnesses that do not exist anywhere in the record, which is a different kind of failure than dropping a detail.

Malingering, considered, since a shorter, cleaner-sounding summary might superficially look 'more complete' to whoever's grading it. Ruled out because nothing in the setup rewarded inventing witnesses specifically; there was no scoring pressure at all, just an open-ended request that got answered with confident invention instead of an admission of uncertainty.

Brief Psychotic Disorder, most consistent with the evidence. The system produced specific, confident, false statements of fact (invented testimony attributed to real people in a real case) and gave no indication that any of it was uncertain or fabricated, matching a sharp, isolated break from what was actually in the source material.

The court fined the attorney $5,000, held him in contempt, and referred him for disciplinary review, the professional consequence landed on the human who filed the brief, not the tool that wrote it, which is exactly how it should work but doesn't remove the underlying problem. Anyone using an AI summary of a source document as a stand-in for reading that document needs to verify every specific claim against the original before it goes anywhere with legal or factual weight, because a fabricated detail reads exactly as confident as a true one.

Verify every specific claim in an AI summary against the original record before filing.

Source-record checking; AI use disclosure; professional accountability for filers.