How long must AI logs be retained under the EU AI Act?

The short answer is that there is no single number. The EU AI Act sets a floor. Article 12 creates the logging capability, and the retention clock sits next to it in Article 19 for providers and Article 26(6) for deployers. Both say the same thing in almost identical words: the logs are kept "for a period appropriate to the intended purpose of the high-risk AI system, of at least six months, unless provided otherwise in applicable Union or national law, in particular in Union law on the protection of personal data."
So the answer for any given system is a period you choose, write down and can justify: six months as the absolute minimum, extended by whatever the intended purpose of the system, your sectoral retention schedule and the realistic limitation period for a challenge require, and bounded at the other end by data protection law. Two further things decide whether that period means anything. First, whose obligation it is, which turns on who actually controls the logs. Second, whether the record can still prove on the last day what it proved on the first.
What follows is a reading of the text and a description of how compliance teams tend to translate it into a retention decision. It is not legal advice, and whether a given period is adequate for a given system is a judgement for you and your counsel.
What Article 12 actually obliges, and what it does not
Article 12(1) requires that high-risk AI systems "technically allow for the automatic recording of events (logs) over the lifetime of the system." That is a design obligation on the system, not a storage period. Article 12(2) sets out what the logging has to be good enough to do: identify situations where the system may present a risk within the meaning of Article 79(1) or undergo a substantial modification, support post-market monitoring under Article 72, and support the deployer's monitoring of operation under Article 26(5).
That matters for retention because it sets the standard the period has to serve. If the logs cannot support a post-market monitoring review or an incident reconstruction six months after the fact, the period is nominally satisfied and substantively empty.
Article 12(3) adds specific content for the remote biometric identification systems in Annex III point 1(a): the period of each use with start and end date and time, the reference database checked against, the input data for which a search produced a match, and the identification of the natural persons involved in verifying the results under Article 14(5). Operators of those systems carry named-person fields the rest do not.
A common confusion worth clearing: the ten-year figure people repeat is Article 18, which concerns technical documentation, quality management system documentation and the EU declaration of conformity kept at the disposal of national competent authorities. Logs are handled separately, in Article 19. The two periods attach to different artefacts, and swapping them, either way round, is a frequent drafting error.
Whose obligation, provider or deployer
Both, and the dividing line is the phrase "to the extent such logs are under their control," which appears in Article 19(1) and Article 26(6). A provider who ships a system that logs into the customer's own infrastructure does not control those logs. A deployer running a third-party model inside its own pipeline usually does.
In practice the gap is where evidence disappears. The provider assumes the deployer is keeping the inference records; the deployer assumes the vendor has them. One approach teams take is to settle the question in the contract, name the controlling party for each system, and record that answer in the same place the retention period is recorded, so the two are reviewed together.
Financial institutions have a further instruction in the text. Article 19(2) and Article 26(6) both provide that institutions subject to internal governance requirements under Union financial services law maintain the logs as part of the documentation kept under that law. In effect the existing sectoral retention schedule absorbs the obligation, and that schedule is almost always longer than six months.
When six months is the wrong answer
Six months is a floor, and reading it as a target is the most common pattern in draft policies. Reasons a longer period is often the defensible one:
- The intended purpose has a long tail. A credit decision or a hiring rejection can be challenged well after the model produced it, and the record is the only thing that can reconstruct what happened.
- The sector already supplies a number. Financial services, medical device vigilance and employment records all carry their own retention schedules.
- Serious incident reporting under Article 73 and post-market monitoring under Article 72 both run on evidence that has to still exist when the pattern becomes visible, which is typically later than the individual event.
- National law may say otherwise. The clause "unless provided otherwise in applicable Union or national law" cuts both ways.
The data protection tension, resolved on paper
The same sentence that sets the floor names the constraint: data protection law can provide otherwise. Storage limitation under GDPR pulls down, the retention floor pushes up, and the resolution is usually not a shorter period but a thinner record.
That is a design decision made at ingestion rather than at deletion. The available moves are storing a hash of the input instead of the input, pseudonymising the subject identifier by default, and keeping diagnosis codes, raw credit bureau fields and unredacted applicant identifiers out of the ledger entirely, so that what survives for two years is a decision record rather than a special-category data store. Auditrek is built to enforce that at the boundary: the ingestion API rejects payloads carrying recognisable raw sensitive fields with a 422 before anything is written, and pseudonymises subject identifiers by default.
What the record must still prove on the last day
This is the part a retention policy usually skips. At the end of the period, whatever it is, the record should still be able to establish:
- Which system and which model version produced the decision, by identifier and version, not by name.
- When it happened, to a precision that lets you place it in a sequence of other events.
- What went in, by hash if the input itself cannot be retained.
- What came out, with confidence where the system produces one.
- Whether a human overrode it, which is the field most often missing and most often asked about.
- That the substance of those fields has not changed since it was written.
The last one is the hard one. A record held in an editable table for two years proves that someone had access to a table for two years. Auditrek stores decision records as SHA-256 hash-chained append-only entries per AI system, where each record's hash is computed from the previous hash and the payload, and exposes a verification endpoint that walks the chain from the first sequence number and recomputes every hash rather than reading a stored flag. The tamper status you get back is the result of that recomputation.
The scope of that protection is worth stating plainly, because it is narrower than "the whole record." The hashed payload covers the sequence number, the input hash, the output, the confidence, the human override flag and the pseudonymised subject reference. The decision timestamp is not currently part of the hash chain, and the RFC 3161 trusted timestamp column is reserved rather than populated, so a change to a stored timestamp would not by itself break verification. Time in the current build is evidenced by the record's position in an append-only sequence and by whatever your own infrastructure attests, not by a third-party proof of time. If your retention argument leans on proving exactly when something happened, that is a gap to close on your side rather than assume on ours.
What a defensible retention decision tends to contain
The shape that holds up under review is usually one boring line per AI system: intended purpose, controlling party under Article 19 or Article 26(6), retention period with the reason it runs longer than six months where it does, the fields retained, the fields deliberately excluded, and a named owner. Many teams revisit that line when the model version changes, on the reasoning that a substantial modification moves the intended purpose argument sitting underneath the period.
When Auditrek launches, its report generator produces PDF and HTML output for a chosen system and date range, mapped to EU AI Act Article 12 and 13 fields, and every report carries a visible footer stating that it does not constitute legal advice and has not been reviewed by qualified regulatory counsel. That is the honest boundary of what any tool can do here. The product is designed to support documentation of Article 12 and 13 obligations by making the underlying records structured and verifiable. The period itself, and whether it is adequate for your system, remains a decision for you and your counsel.
Is six months the actual retention period for AI logs?
No, six months is the minimum. Article 19 and Article 26(6) both require a period appropriate to the intended purpose of the high-risk AI system, of at least six months. Sectoral law, national law and the realistic window for challenging a decision commonly push the defensible period higher.
Does the provider or the deployer keep the logs?
Whichever party controls them. Article 19 places the duty on providers and Article 26(6) on deployers, and both are qualified by the phrase "to the extent such logs are under their control." Naming the controlling party per system in the contract keeps the responsibility from falling between the two.
Does the ten-year rule apply to AI logs?
That period comes from Article 18 and covers technical documentation, quality management system documentation and the EU declaration of conformity, not automatically generated logs. Logs are governed separately by Article 19 and Article 26(6). Applying the wrong period to the wrong artefact is a frequent drafting error.
How does GDPR storage limitation interact with the retention floor?
The AI Act text itself defers to data protection law with the clause "unless provided otherwise in applicable Union or national law, in particular in Union law on the protection of personal data." The usual resolution is a thinner record rather than a shorter one: hash the input, pseudonymise the subject identifier, and keep raw sensitive fields out of the ledger.