What the logs actually contain

Across financial services, healthcare, legal, and government, the current state of llm auditability is: timestamps and token counts. Sometimes the prompt. Almost never a cryptographically verifiable record that the stored log matches what was actually sent and received. The logs are mutable. Anyone with database write access can edit them. The model vendor's own logs — if they exist — are controlled by the vendor.

The case that made this concrete

In 2023, a european bank deployed a large language model to assist loan officers with credit risk summaries. Eighteen months later, a customer filed a discrimination complaint. The compliance team went looking for the interaction log. They found a timestamp, a user id, and a token count. The actual prompt and response had been deleted after 30 days per their data retention policy. That bank is not unusual. It is the norm.

What the regulations are starting to require

The eu ai act, entering full enforcement in 2026, requires that high-risk ai systems maintain logging sufficient for post-hoc monitoring. The ftc has already brought actions against companies whose ai outputs could not be reconstructed. The nist ai risk management framework lists traceability as a core governance requirement. None of these specify what adequate logging looks like technically. But every legal reading points in the same direction: a mutable database row is not going to be enough.