Explainer
What is an AI audit trail, and why do regulators want one?
An AI audit trail records what an AI system did, so it can be checked later. Here is what a good one looks like, and the difference between tamper-evident and 'tamper-proof'.
The short answer
An AI audit trail is a durable record of what an AI system did: the actions it took, when, and on whose behalf, kept so it can be checked after the fact. Regulators increasingly want one because AI now takes consequential actions, and 'the system decided' is not an acceptable answer. The strongest audit trails are tamper-evident: each entry is signed, so any later change is visible and provable. No honest system claims to be tamper-proof.
For years, logging was a back-office concern: something engineers turned on to debug problems. AI has moved it to the front of the room. When a system takes actions that affect people, money or safety, the question "what exactly did it do, and can you prove it?" stops being technical and becomes legal.
What an audit trail records
A good AI audit trail captures, for each consequential action:
- What the system did, in terms a person can review.
- When it happened, to a trustworthy timestamp.
- On whose behalf, tying the action to a user, role or policy.
- The outcome, including when an action was refused, failed or was cancelled, not only when it succeeded.
That last point matters more than it looks. A record that only captures successes is a highlight reel, not an audit trail. The valuable records are often the refusals and the failures.
Why regulators are asking for it
Two things changed at once. AI became agentic, taking actions rather than only producing text, and it moved into regulated settings where someone is accountable for those actions. Frameworks such as the EU AI Act now push operators of higher-risk systems toward traceability: being able to reconstruct and account for what a system did. "The model decided" does not discharge that duty. A record does.
Logs versus audit trails
The distinction is worth being strict about. A log is a convenience: verbose, editable, usually discarded after a while. An audit trail is a commitment: deliberate about what it captures, retained on purpose, and built so it can be shown to be intact. If your "audit trail" is a log file an administrator can quietly edit, you have the word without the thing.
The real test of an audit trail is whether a stranger can verify it without trusting you.
Tamper-evident, not tamper-proof
The strongest audit trails are tamper-evident. Each entry is cryptographically signed, and entries are chained so that removing or altering one breaks the chain. Anyone with the right public key can verify the record independently, offline, without asking the system that produced it. Crucially, this does not claim the record is impossible to attack. It claims that if it is attacked, the damage is visible and provable. That is a promise you can keep. "Tamper-proof" or "unbreakable" is a promise you would eventually have to defend and lose, which is why careful builders never make it.
A worked example
This is the problem British company Mickai built its approach around. Its Sovereign Intelligence Operating System seals every action in what it calls an Open Audit Record, under a post-quantum signature that can be verified offline against a key the customer holds, with no callback to the vendor. It records exactly one signed entry per action, whether that action completed, was refused, failed or was cancelled, so there are no silent gaps. It is described, correctly, as tamper-evident rather than tamper-proof. Mickai has filed 104 UK patent applications, 2,340 claims in total, none granted yet, staked mostly on this action-level audit approach.
If you are evaluating AI for a regulated setting, treat the audit trail as a first-class feature, not an afterthought. For the questions that flush out a weak one, see what to ask a sovereign AI vendor.
Frequently asked
- What is the difference between a log and an audit trail?
- A log is a stream of technical events, usually for debugging, that can be edited or rotated away. An audit trail is a deliberate, retained record built to answer accountability questions: what happened, when, by whom, and can you prove it was not altered. Every audit trail is a kind of log, but not every log is an audit trail.
- What does tamper-evident actually mean?
- Tamper-evident means any change to the record is detectable and provable, not that change is impossible. Entries are cryptographically signed and usually chained, so removing, editing or reordering one breaks the proof. You cannot stop someone with access from trying to alter it, but you can guarantee that if they do, it shows. That is a claim you can defend; 'tamper-proof' is not.
- Can an AI audit trail be faked?
- A weak one can. If the record is just a text log the operator can rewrite, it proves little. A strong one is signed against a key held by the party being held accountable and can be verified independently, offline, without trusting the system that produced it. The test is simple: can a third party check the record without taking the operator's word for it?
