Audit Logging (AI)

In traditional software, an audit log records who did what and when. In an AI system, the logging surface is dramatically more complex: the "what" is a probabilistic generation from a model with billions of parameters, the "why" is not directly inspectable from the log entry, and the "with what data" spans a dynamic retrieval process that may have incorporated hundreds of documents. AI audit logging must therefore capture more context than traditional application logging to be genuinely useful for compliance and incident response.

A complete AI audit log captures events across four layers: (1) **Request layer** — the user identity, timestamp, input prompt or query, and session context; (2) **Retrieval layer** (for RAG systems) — the retrieved document IDs, similarity scores, and access control decisions made during retrieval; (3) **Model layer** — the model and version invoked, relevant configuration parameters, and the full output including any intermediate reasoning steps (chain-of-thought); (4) **Action layer** (for agents) — every tool call, external API request, code execution, and file system operation with its inputs, outputs, and success/failure status.

The retention, storage, and integrity requirements for AI audit logs differ significantly by use case and jurisdiction. GDPR Article 22 requires records supporting automated decision-making explanations. The EU AI Act requires audit trail retention for high-risk AI. US financial regulation (SR 11-7) requires model governance documentation. Healthcare AI operating under FDA guidance requires lifecycle records. Enterprise AI teams should implement log integrity controls (cryptographic hashing, append-only storage, immutable cloud storage buckets) to ensure audit logs are admissible as evidence and cannot be retroactively modified.