Article 12 logging requirements: what to record, for how long, and how to prove it
Article 12 is four paragraphs long and most engineering teams get it half right: they log the model call, then stop. What the Act asks for is a record of events that can reconstruct a decision and show whether the system drifted — which is a different design.
Article 12 of Regulation (EU) 2024/1689 is short. High-risk AI systems must be able to record events automatically over their lifetime, and the logging must make it possible to identify risks, support post-market monitoring and monitor how the system is operated. Everything else in this article is about turning that sentence into something an auditor can accept.
What Article 12 actually requires
Paragraph 1 is a design requirement, not an operations requirement: the system shall technically allow for the automatic recording of events over its lifetime. "Technically allow" means logging is a capability of the system, on by default for high-risk operation, not something a customer switches on after an incident.
Paragraph 2 says the logs must be usable for three purposes: identifying situations that may present a risk or lead to a substantial modification, facilitating post-market monitoring under Article 72, and monitoring the operation of the system by the deployer under Article 26(5). A log that only stores the prompt and the response supports none of the three.
For systems in Annex III point 1(a) — remote biometric identification — the second subparagraph of paragraph 2 sets four mandatory elements: the period of each use, the reference database against which input data was checked, the input data that produced a match, and the identity of the natural persons involved in verifying the result under Article 14(5). If you build biometric identification, those four fields are not optional.
Providers and deployers have different duties
| Role | Article | Duty |
|---|---|---|
| Provider | Art. 12 | Build logging into the system; document it in the technical file (Annex IV, point 1(c)) |
| Provider | Art. 19 | Keep the logs under your control for at least six months, longer if sector law requires |
| Deployer | Art. 26(5) | Monitor the system in operation and inform the provider about risks |
| Deployer | Art. 26(6) | Keep logs under your control for at least six months |
| Deployer | Art. 26(6) | Retain logs where they are not under the provider's control |
In practice most small deployments are deployers: you bought or subscribed to a model or an application and you run it on your own data. That means the six-month floor in Article 26(6) is your duty, and the provider's logging feature is only useful to you if you can export it and keep it.
Many AI vendors keep logs for you and show them in a dashboard. That satisfies nobody if the retention period is 30 days, the export is a screenshot, or the account is closed at the end of the contract. Ask for a machine-readable export with timestamps before you sign.
The minimum record: six fields
Article 12 does not prescribe a schema, but the three purposes it names do. A record that cannot answer these six questions cannot support a risk assessment, a monitoring report or an oversight review.
| Field | Why the Act needs it | Common mistake |
|---|---|---|
| When (UTC, trustworthy clock) | Period of use, ordering, retention | Local timestamps without offset; clocks drifting between services |
| Who acted (user or service identity) | Oversight traceability, access review | A generic "api" actor for every call |
| What happened (event type) | Distinguishing inference, approval, error, data access | One event type for everything, so approvals cannot be separated |
| Which model and version | Substantial modification, post-market monitoring | Model family only, or the model name at query time rather than at log time |
| Which data categories | Risk assessment, minimisation, DPIA evidence | Storing raw personal data where a category label would do |
| Who approved, if anyone | Article 14 human oversight | An approval flag with no identity and no timestamp |
That is the minimum. Systems that make decisions about people usually need two more: a stable decision or case identifier, so the log can be joined to the business record, and a correlation ID across services, so a chain of calls can be reconstructed as one event.
A schema that survives an audit
{
"seq": 1042,
"occurred_at": "2026-11-16T09:12:44.001Z",
"received_at": "2026-11-16T09:12:44.180Z",
"actor": "recruiter@example.com",
"event_type": "inference",
"model": { "provider": "openai", "name": "gpt-4.1-mini", "version": "2026-08-01" },
"data_categories": ["employment", "email"],
"input_summary": "Rank 42 applicants for warehouse role",
"output_summary": "Shortlist: 12, 27, 31",
"decision_id": "case_88213",
"approval_id": "ap_7f3c",
"prev_hash": "3c5a7e9f...",
"hash": "9f2c7a1e..."
}Two design choices in that schema matter more than the field names. First, the actor is a person or a named service, never a shared API key. Second, the record commits to the previous record's hash: changing a historical row breaks every hash after it, which is what turns a log into evidence.
Retention: six months is a floor, not a strategy
Articles 19 and 26(6) both say at least six months, "unless provided otherwise in applicable Union or national law, in particular in Union or national law relating to the protection of personal data". That cuts both ways.
- Sector rules can extend it. Employment, financial services and health regulation often require years, not months.
- GDPR can shorten it. Logs contain personal data; storage limitation applies, and a six-month log of every applicant you rejected needs a lawful basis and a documented retention rule.
- Litigation can freeze it. Once a claim or an investigation starts, deletion becomes spoliation. A retention policy needs a legal hold switch.
- Contracts can raise it. Enterprise customers increasingly write AI Act retention into the DPA, and six months will not satisfy a three-year audit cycle.
Write the rule down, apply it automatically, and log the deletions. An auditor who finds logs older than your stated retention period will ask why; an auditor who finds unexplained gaps will ask harder questions.
Integrity: the part Article 12 implies
The Act does not use the word "hash". It does require logs that support monitoring, incident investigation and, in practice, demonstration to a market surveillance authority. A mutable database table with an admin account that can run DELETE satisfies none of that, and every serious auditor knows it.
The cheapest credible answer is a hash chain: each record includes the hash of the previous record plus its own content, and a periodic digest of the chain head is stored somewhere the application cannot rewrite. Add a signature over the periodic digest and you have a record that can be verified by a third party without access to your systems.
TraceSeal ingests OpenAI, Azure OpenAI, Ollama or custom events, masks personal data before storage, and chains every record. One click produces a signed pack with the chain head and Merkle root.
Open the verification pageThe five mistakes that fail an audit
- 1Logging outputs only. Without the input and the actor, you cannot reconstruct why a decision was made.
- 2No model version. When the provider silently updates the model, your evaluation no longer describes what ran.
- 3Approvals outside the log. A tick in a ticketing system that is not joined to the event is not oversight evidence.
- 4No retention rule. Either you keep everything forever (GDPR problem) or nothing (AI Act problem). Both are findings.
- 5No way to verify. If the only proof that the file is unchanged is your word, the file is a claim, not evidence.
How this connects to the rest of the Act
Article 12 is the plumbing for four other obligations: Article 14 (human oversight needs a record of who reviewed what), Article 72 (post-market monitoring needs a baseline to compare against), Article 73 (serious incident reporting needs a timeline) and Article 17 (the quality management system needs documented processes, including logging). A logging design that only serves engineering observability will leave all four short.
Under Regulation (EU) 2026/1744 the Chapter III deadline for Annex III systems moved to 2 December 2027. Article 12 is part of Chapter III, so the date moved with it — but Article 50 transparency duties have applied since 2 August 2026, and buyers are already asking for the file.
Frequently asked
Does Article 12 require logs to be immutable?
It requires automatic recording that supports risk identification, post-market monitoring and operational monitoring, and Article 19 requires keeping those logs. It does not prescribe a technical mechanism. Tamper-evidence (hash chain, append-only storage, WORM) is the practical way to show the logs were not altered.
How long do we keep AI Act logs?
At least six months under Article 19 for providers and Article 26(6) for deployers, unless Union or national law says otherwise. Sector rules, contractual audit rights and GDPR storage limitation all push in different directions, so the retention period should be written in the quality management system with a legal basis for it.
Do we have to log prompts and responses verbatim?
Not necessarily. The Act needs enough to identify risks and reconstruct decisions; storing raw personal data increases your GDPR exposure. A masked summary plus data-category labels plus a hash of the original is usually a better trade-off than full payload retention.
We use a vendor API. Who is responsible for the logs?
Both, in different ways. The provider must build logging into the system; the deployer must keep the logs under its control for at least six months. If your vendor retains logs for 30 days and offers no export, you are the one with the gap.
Sources
Keep reading
An AI Act evidence pack is the fixed record an operator shows when asked how a high-risk AI system is governed: classification, controls, logs, human oversight and a way to prove nothing was edited.
Which AI systems Annex III makes high-risk, how the Article 6(3) filter works, the full obligation checklist under Articles 9–15, 17, 43 and 72, and the timeline after the AI Digital Omnibus.