AI audit logs

Audit logs for an AI system are a durable record of what the AI was asked, what it did and which policy fired. They let you reconstruct an incident and show an auditor that your rules were applied.

What to log

  • User, session and time.
  • The input and the output, or a redacted form of each.
  • Every tool call, with its arguments and result.
  • Each policy verdict and the action taken, such as blocked, flagged or held for approval.
  • The versions of the model, the prompt and the policy in force.
  • Human decisions: an approval, a review.

An audit log differs from a trace in purpose. A trace is for debugging. An audit log is evidence for AI compliance, so entries are appended and never edited.

Retention and redaction

Prompts and outputs often hold personal data and secrets, so the log is itself sensitive. Limit who can read it, and settle two things before turning it on:

  • Retention. Keep entries as long as the rules you answer to require, and no longer.
  • Redaction. Mask personally identifiable information (PII) before it is written. See PII redaction.

With White Circle

Analyze has audit logs for compliance and incident review. Logging is opt-in and configurable, with control over redaction and retention, and conversations are not stored unless logging is turned on.

Train your own models

Post-train any Hugging Face model with Halo. Up to 4× faster than TRL