New Mexico appeal exposes fabricated testimony in an unchecked AI-assisted brief

Public defenders, appellate supervisors and court IT teams need to reduce drafting burden without losing control of the record. Ask how transcripts enter the workflow, who checks witness identities, and whether reviewers can see the exact cited passage. A bounded engagement could examine one appeal-preparation process and build a verification exercise using authorized historical material. The value hypothesis is identifying preventable review failures and quantifying the effort required for safe assistance. Avoid promising better appeal results, zero hallucinations or savings from this incident. Suitability depends on local practice and the team's capacity to perform substantive review, including under deadline pressure.

Fit is an isolated drafting workspace with document-level permissions and source-linked assertions. Prerequisites include reliable transcripts, stable passage identifiers and an attorney-adjudicated reference set. Test nonexistent names, misattribution, contradictory passages and unsupported quotations using the proposed current model. Compare assisted reviewers with the existing process and measure both missed errors and correction effort. Retrieval can support checking but does not guarantee fidelity. Cloud, local and hybrid choices require separate privacy and operational assessment; this story establishes no product compliance status. Keep generation separate from case-system updates and filing credentials, and log model changes for repeat evaluation.

An appellate practice lead should own the process with records staff, training personnel and IT. Map transcription, draft creation, review and signature; establish a conventional baseline and run a reversible shadow pilot. Dependencies include protected review time and lawful access to the record. Train staff to verify facts as well as legal references, and provide accessible source-navigation instructions.

Proposed acceptance

every factual assertion in a sampled final draft has a checked source location, and any fabricated witness stops release pending correction. These are proposed controls, not observed results. Track reviewer workload and abandoned drafts; preserve a manual route when validation cannot be completed.