What does an AI audit trail need to contain?

Enough for someone else to check it: which key called which model, when, what went in and what came out, in a record that cannot be edited afterwards without it showing. Content hashes rather than content keep the text private while still committing to it, and chaining each entry to the previous one makes a missing entry detectable.

The four parts

Identity — whose key. Provenance — which model. Time. And a commitment to the content, which a salted hash gives you without publishing the content itself.

Why chaining matters

An entry that points at the one before it turns a deletion into a visible gap. Without it, the absence of a record and the removal of a record look identical.

Who checks it

An internal auditor, a client, a regulator, or the other side of a dispute. A trail only one party can read is a trail that party is asking to be believed about.

Open the console Read the docs

More answers: How do I prove which AI model answered? · Is there an OpenAI-compatible API with an audit trail? · Can I buy AI inference billed in dirhams? · Which LLM APIs support Arabic properly?