← LearnEU AI Act

AI provenance and the EU AI Act: the evidence you actually need

Signed outputs, an append-only log, and a crypto inventory produce evidence that supports your obligations — not a certification.

What the Act pushes toward

The EU AI Act, whose obligations for general-purpose and high-risk systems become broadly applicable in 2026, moves organizations toward AI they can attribute, document, and audit. In practice that shows up as transparency to users, record-keeping and logging, and technical documentation of the system.

None of that is satisfied by a marketing claim. An assessor wants artifacts: evidence that a given output really came from your system, an auditable record that it hasn’t been quietly changed, and documentation of the components and cryptography involved.

How cryptographic provenance produces that evidence

A signed receipt on each answer establishes attribution and integrity — the output came from this system, unaltered — which is exactly the kind of tamper-evident record transparency and record-keeping duties call for.

An append-only transparency log (built like Certificate Transparency, publishing only hashes, never content) gives you a record a past output cannot be silently rewritten in. And a Cryptographic Bill of Materials documents every primitive in use, mapped to its standard — the kind of technical documentation an assessor can actually check.

The honest line: evidence, not compliance

This is the part to be careful about. Cryptographic provenance SUPPORTS these obligations by producing evidence for them. It does not make you “compliant,” and it is not a certification or a declaration of conformity.

Which obligations actually apply — and to whom — depends on your system’s risk classification under the Act, which only you and qualified counsel can assess. Any tool that hands you a “you’re compliant” verdict is overreaching; the right output is factual evidence you can put in front of an assessor, plus a clear note that the legal determination is yours to make.

What this looks like in practice

The concrete deliverable is a portable, signed evidence artifact: the answer, its post-quantum signatures and the public keys to check them, and the obligation mapping — an object your counterparty or assessor can re-verify themselves, offline, years later, without trusting your servers.

That turns an abstract obligation (“maintain documentation you can demonstrate to third parties”) into a file you can hand over and they can check — which is the whole point of provenance.

Generate an AI Provenance Record

Factual evidence that supports your obligations — not legal advice, and not a declaration of conformity.

Verification attests an answer’s origin and integrity, not its factual accuracy. Algorithm names denote the public standards the primitives are based on (ML-DSA-87 / FIPS 204, ML-KEM-1024 / FIPS 203; Falcon / FN-DSA, FIPS 206 forthcoming), not a FIPS-140 / CMVP validation.