Article 12 is the record-keeping obligation at the centre of the EU AI Act's traceability regime. It requires high-risk AI systems to record events automatically over their lifetime. What it does not settle is whether the resulting logs can be relied upon by anyone other than the operator that holds them.
The obligation in plain terms
Article 12 requires that high-risk AI systems technically allow the automatic recording of events over the lifetime of the system. The logging must be appropriate to the intended purpose and must enable the identification of situations where the system may present a risk within the meaning of Article 79, or where it undergoes a substantial modification. It also underpins post-market monitoring under Article 72.
For certain high-risk systems, Article 12 sets out minimum categories of recorded information, including the period of each use, the reference data against which inputs were checked, the inputs for which a search produced a match, and the identity of the natural persons involved in verifying results.
Who it applies to
- Providers of high-risk AI systems, who must design the system so that logging happens automatically, and who retain the logs they generate.
- Deployers of high-risk AI systems, who under Article 26 keep the logs generated by systems under their control.
Whether a system is high-risk is determined by Article 6 and Annex III together with the provider's own assessment. NexArt does not perform that classification.
Retention under Articles 19 and 20
Article 19 requires providers to keep the logs automatically generated by their high-risk AI systems for a period appropriate to the intended purpose, and at least six months unless Union or national law provides otherwise. In practice this often collides with ordinary log hygiene: application logs are rotated, sampled, and downsampled long before six months elapse.
The reliability question
Article 12 says events must be recorded. It does not say how a supervisory authority should satisfy itself that what it is shown is what was recorded. That gap matters, because ordinary logs have three properties that weaken them under scrutiny:
- They are mutable. Anyone with database or pipeline access can change an entry, and the entry itself does not distinguish an original from a later edit.
- They are operator-owned. Their credibility depends on trusting the party whose conduct is under examination.
- They are not reproducible. Re-running a prompt does not recreate the earlier output, because model versions, parameters, and providers change.
This is the difference between governance and verification. Governance describes the controls you intend to operate; verification establishes what actually executed. See AI governance vs AI verification.
How Certified Execution Records support Article 12
- Record at the decision boundary. One SDK call at each AI decision point produces a Certified Execution Record covering input, output, model context, parameters, and optional identity binding.
- Seal it. Protected fields are canonically serialised and hashed. Under confidential execution, sensitive values are stored as keyed commitments, supporting data minimisation.
- Attest independently. The NexArt attestation node returns an Ed25519 signature and an RFC 3161 timestamp issued via DigiCert's public timestamp authority.
- Retain as portable JSON. Records archive alongside your existing Article 12 logs and Article 11 technical documentation, not instead of them.
- Verify without the operator. A deployer, auditor, or supervisory authority holding the record and the published public key can check integrity offline.
Implementation checklist
- Identify the decision points in the system that materially affect an outcome for a person.
- Certify each of those executions rather than sampling them.
- Bind records to a case, subject, or account identifier so a specific run can be located later.
- Group multi-step agent flows into a Project Bundle so ordering across steps is covered.
- Archive records for at least the Article 19 minimum, independently of log rotation policy.
- Document the verification procedure so a third party can follow it without your assistance.
Scope and limits
- NexArt does not classify AI systems as high-risk.
- NexArt does not perform conformity assessments and is not a notified body.
- NexArt is not itself certified under the EU AI Act, ISO/IEC 42001, or SOC 2.
- Verification proves integrity, not correctness. A record shows what executed. It does not show that the decision was accurate, fair, or lawful.
- Article 12 sits inside a wider governance regime. Risk management, human oversight, and quality management obligations are unaffected by the existence of verifiable records.
See Non-goals for the full boundary of what NexArt does not do, and EU AI Act audit trail for the mapping across Articles 11, 13, 20, and 72.
Frequently asked questions
- What does EU AI Act Article 12 require?
- Article 12 requires high-risk AI systems to technically allow the automatic recording of events, logs, over the lifetime of the system. The logging must enable identification of situations where the system may present a risk or undergo a substantial modification, and must support post-market monitoring under Article 72.
- How long must Article 12 logs be kept?
- Retention is governed by Article 19 and Article 20. Providers keep automatically generated logs for a period appropriate to the intended purpose of the system, and at least six months unless applicable Union or national law provides otherwise. Certified Execution Records are portable JSON and can be archived alongside existing retention systems.
- Do Article 12 logs have to be tamper-evident?
- The Regulation does not prescribe a specific technical control. It does require the logs to be reliable enough for supervisory authorities and deployers to reconstruct how the system operated. Cryptographic sealing is a recognised technical means of establishing that a record has not changed since it was created.
- Does NexArt make our high-risk AI system compliant with Article 12?
- No. NexArt produces verifiable execution evidence that supports Article 12 logging integrity and traceability. It is not a governance platform, a conformity assessment, or a compliance certification. Compliance determinations remain with the provider, the deployer, their notified body, and the competent supervisory authority.
- Can Article 12 logging be satisfied without storing personal data?
- Confidential execution is the default in NexArt. Inputs, outputs, and identity fields are stored as keyed cryptographic commitments rather than raw values, so records remain verifiable while supporting data minimisation. Specific values can be revealed selectively and checked against the commitment.
Ready to see execution evidence in practice?
Start producing Certified Execution Records in under five minutes, or verify an existing record independently.