The Witness Never Existed. The Brief Said Otherwise.

AI hallucination becomes an institutional failure when fabricated output is promoted into evidence.

In a September 9, 2026 dispositional order, the Supreme Court of New Mexico held attorney Stephen D. Aarons in direct contempt after an appellate brief in a murder appeal contained false factual material produced with ChatGPT. The court described fabricated testimony attributed to witnesses who did not exist, false testimony attributed to actual witnesses, and inaccurate descriptions of legal authority. Aarons acknowledged that he had not verified the factual assertions or legal authorities before signing and filing the brief.

The court struck the briefing, removed him from the appeal, barred him from appearing before the court pending disciplinary proceedings, referred the matter to the Disciplinary Board, and imposed a $5,000 sanction.

The failure was not only generation

The source trial record was real. The fabricated testimony was not. The AI output introduced unsupported assertions between Source and Evidence, but the institutional consequence occurred when those assertions passed through human review and entered an authoritative filing.

GovKM therefore maps the incident as a compounded continuity break: Source → Evidence failed because synthetic claims were not reconciled to the source record, and Evidence → Authority failed because the signing attorney promoted unverified output into an institutional artifact.

Human review must be an evidentiary control

It is not enough for a human to remain nominally “in the loop.” If the human authority does not verify whether a generated claim is supported by the source evidence, the review step becomes procedural rather than substantive.

That distinction is central to trustworthy AI governance. Generated material can be useful analysis, candidate context, or a research aid without being evidence. The transition from generated output to evidence should be observable, governed, and attributable to the authority that accepted it.

GovKM interprets this case as evidence for a simple rule: AI output should not become institutional evidence merely because a human copied, signed, or submitted it. Promotion must include source reconciliation.

Source

Supreme Court of New Mexico, “State v. Sandoval / In re Stephen D. Aarons — Dispositional Order of Direct Contempt,” No. S-1-SC-40845, September 9, 2026.

Published court order