Audit Log

An audit log is a record that captures, in order, the actions an AI system took. Think of it like an airplane's black box: nobody checks it in ordinary times, but when something goes wrong, it becomes nearly the only evidence for reconstructing what happened. It records which functions the AI called, what values it passed in, and which external systems it connected to.

This has become important because AI systems have shifted toward making their own decisions and carrying out tasks through multiple steps. When a human clicks through a screen, a screen recording is enough. But when an AI handles the intermediate steps on its own, humans never actually see that process. Saving only the question and the answer isn't enough anymore. To reconstruct the cause of a problem later, you need a record of what was called and where it connected to along the way.

This is also why the term comes up so often in regulatory discussions. AI systems classified as high-risk are frequently required to retain this kind of log for a set period. The idea is that both the company and regulators need a way to verify what actually happened if an incident occurs.

In the article, auditable execution logs appear as one of the items the EU may require of OpenAI and Anthropic. The context involves logging records at the level of each tool call and setting a retention period for how long they must be kept. A common misconception is that saving only the questions and answers shown in a chat window is enough — but in a structure where AI calls tools and connects to external systems, the intermediate steps also need to be recorded to reconstruct how an incident occurred.