What should a healthcare AI audit trail capture?

The record should answer four questions: what evidence entered the workflow, what the AI produced, what an authorized person did, and what was released or written downstream. For each clinically meaningful event, retain the source or record reference and version; active workflow, rule, prompt, and model version; the AI output; material edits; reviewer identity and role; approval, rejection, escalation, or override; timestamp; and the result of any export or system writeback. Preserve failed attempts and exceptions as well as successful actions so the institution can reconstruct the path to the final record.

An audit trail is not a reason to duplicate every patient datum. The institution should minimize captured data, restrict access by role, protect integrity, define retention, and test whether authorized reviewers can retrieve the evidence when needed. The log should distinguish an AI draft from a human-approved record and should not allow the system that produced an output to certify its own correctness. An audit trail supports investigation and accountability; it does not by itself prove that a clinical result was correct or safe.