From 2 December 2027 (for systems already in use, by 2 August 2030 at the latest), a municipality using a high-risk AI system keeps the logs that the system generates automatically and that are under its control for a period appropriate to the intended purpose of the system, of at least six months, unless other Union or national law, in particular data protection law, provides otherwise (Article 26(6) AI Act). Does the GDPR allow this? In principle, yes. Six months is a statutory floor, not a free pass: anything the municipality keeps for longer it must justify under the GDPR itself, and it must secure the logs as the personal data they are.
Suppose the Work and Income department uses an AI system that assesses applications for social assistance and gives the caseworker a recommendation for each application. Every recommendation, every consultation and every time a caseworker follows or sets aside the recommendation leaves a trace. The question the privacy officer then gets: how long do we keep that, who may access it, and what goes into the record of processing activities?
In short
A high-risk AI system must technically allow for the automatic recording of events over its lifetime (Article 12(1)). Those logs serve three purposes: identifying risks and substantial modifications, post-market monitoring by the provider, and monitoring of the operation by the deployer (Article 12(2)). The provider keeps the logs under its control (Article 19), the deployer the logs under its own (Article 26(6)), in both cases for a period appropriate to the purpose and of at least six months. Financial institutions keep the logs as part of the documentation they already maintain under financial services law. For Annex III systems these obligations apply from 2 December 2027; for systems already in use before then, the transitional rule in Article 111(2) applies. The GDPR continues to apply in full: logs that can be linked to an applicant or an employee are personal data, so storage limitation, security, the right of access and accountability all apply.
This piece goes deeper into the retention period and the GDPR. How logging fits into the wider work of IT and data teams is covered in the AI Act belongs with IT.
What Article 12 requires: logging as a design requirement
Article 12 AI Act is titled "Record-keeping". Paragraph 1 says that high-risk AI systems "shall technically allow for the automatic recording of events (logs) over the lifetime of the system". That is a requirement for the product, and therefore an obligation of the provider. The municipality buying the system does not build the logging itself, but it does need to be able to check that it is there.
Paragraph 2 sets out what the logging has to be good for. Logging capabilities must enable the recording of events relevant for:
- identifying situations that may result in the system presenting a risk within the meaning of Article 79(1), or in a substantial modification;
- facilitating the provider's post-market monitoring (Article 72);
- monitoring the operation of the system by the deployer (Article 26(5)).
For most high-risk systems the Regulation does not give a list of fields. The level of traceability must be "appropriate" to the intended purpose. Only for remote biometric identification systems (Annex III, point 1(a)) does paragraph 3 set a minimum: the period of each use (start and end date and time), the reference database against which input data was checked, the input data for which the search led to a match, and the identification of the natural persons involved in verifying the results. That last element ties in with Article 14(5): for those systems, the deployer may not act on an identification unless it has been separately verified and confirmed by at least two natural persons with the necessary competence, training and authority, save for the exception for law enforcement, migration, border control and asylum where Union or national law considers that requirement disproportionate. The log shows whether that happened.
Paragraph 3 does not apply to the Work and Income system. What is logged then follows from the three purposes in paragraph 2 and from the instructions for use. Those instructions must, where relevant, describe the mechanisms that allow deployers to properly collect, store and interpret the logs (Article 13(3)(f)); see also what Article 13 requires of instructions for use. On our reading, a log for such a system is only useful for the purposes in paragraph 2 if, for each recommendation, it records at least: when the recommendation was given, with which model version, on which input (preferably as a reference to the case file rather than a copy), what the output was, which employee saw the recommendation and whether they followed it or departed from it. That is our practical translation, not a list taken from the Regulation.
Who keeps what: Article 19 and Article 26(6)
The Regulation splits retention between two roles, in each case for the logs "to the extent such logs are under their control".
| Provider | Deployer (the municipality) | |
|---|---|---|
| Legal basis | Article 19(1) and Article 16(e) | Article 26(6), first subparagraph |
| Which logs | The Article 12(1) logs under its control | The automatically generated logs under its control |
| Period | Appropriate to the intended purpose, at least six months | Appropriate to the intended purpose, at least six months |
| Proviso | Unless provided otherwise in applicable Union or national law, in particular data protection law | Same |
| Financial institutions | Keep the logs as part of the documentation kept under financial services law (Article 19(2)) | Same (Article 26(6), second subparagraph) |
| Towards the authority | Gives access to the logs under its control upon a reasoned request (Article 21(2)) | Cooperates with the competent authorities (Article 26(12)) |
"Under their control" is a contract question. If the system runs as a cloud service at the vendor, the logs often sit with the vendor. It then keeps them as provider under Article 19, but the municipality can only meet its own monitoring obligation if it has access. Put in the contract which logs are created, that you can view and export them, how long the vendor keeps them and when it deletes them. Under the GDPR the vendor is often your processor for those logs; if it also keeps them for its own obligations as provider (Articles 19 and 72), it may be a controller for that part. Set out that allocation of roles in the contract and the data processing agreement.
How long, then? Six months is the floor
The AI Act sets a minimum and leaves the rest open. What "appropriate to the intended purpose" means is for the municipality to determine. The GDPR sets no periods, but it does provide the test: personal data must be kept in a form which permits identification of data subjects for no longer than is necessary for the purposes for which they are processed (Article 5(1)(e) GDPR, storage limitation).
For logs, that test is workable if you reason back from what they are for. Three questions help:
- How long can a decision still be challenged? A log showing which recommendation the system gave and what the caseworker did with it is relevant for as long as an applicant can object or ask for an explanation about a decision in which the system played a role. That can justify a period longer than six months, provided you justify it per type of log.
- What does monitoring need? To detect drift or a rising number of deviations you need time series, but not necessarily data that can be traced back to individuals. After a certain period, aggregated or anonymised figures may be enough.
- Does other law apply to the same data? Article 19(1) expressly applies without prejudice to other Union or national law, and both provisions give way where that law provides otherwise. A retention obligation that already applies to the case file therefore does not lapse because of the AI Act.
The European Data Protection Board (EDPB) is explicit about logs in its Guidelines 4/2019 on Data Protection by Design and by Default (version 2.0, adopted on 20 October 2020). Among the elements of storage limitation, it says that controllers shall determine what personal data and length of storage is necessary for back-ups and logs, that they must be able to justify why a period is necessary, and that deletion should preferably be automated (point 82). Once personal data is no longer necessary for the purpose, it shall by default be deleted or anonymised, with systematic procedures for that embedded in the processing (point 53). According to the EDPB, any retention should be objectively justifiable as necessary, in line with the accountability principle (point 52).
One open point. The AI Act says "at least six months, unless provided otherwise" in data protection law. Whether the GDPR could in a specific case require a shorter period has not been settled; we know of no official interpretation by the Commission, the EDPB or the Dutch Data Protection Authority. In practice the safe line is: six months as the floor, longer only with a justification per type of log.
Logs are personal data
A log stating that application 4711 received a negative recommendation on 3 March at 10:14 and that caseworker J. followed it concerns two people: the applicant and the caseworker. Both are identifiable, so that log line contains personal data of both within the meaning of Article 4(1) GDPR.
The Court of Justice worked this out for log files in Pankki S (C-579/21, 22 June 2023). A (since dismissed) employee and customer of a bank wanted to know who had consulted his customer data, when and why. The Court held that information relating to consultation operations carried out on a person's data, concerning the dates and purposes of those operations, is information that person is entitled to obtain under Article 15(1) GDPR. There is in principle no such right as regards the identity of the employees who carried out the consultations under the authority and on the instructions of the controller, unless that information is essential to enable the data subject to exercise the rights conferred by the GDPR effectively, taking into account the rights and freedoms of those employees (point 83). The Court also noted that the log files in that case fell within the records of processing activities under Article 30 GDPR and serve to demonstrate that processing complies with the GDPR (point 67).
For the municipality, that means three things:
- Access requests reach the logs. If an applicant asks for access, according to the Court the dates and purposes of consultations of their data form part of the answer. On our reading, this means the dates and purposes of consultations and processing by the AI system also form part of the answer. Make sure you can retrieve logs per person.
- The employee is a data subject too. Logs of who followed which recommendation are also personal data of caseworkers. Use them to oversee the operation of the system, and decide in advance whether and how they may be used for other purposes, such as staff appraisal.
- Tell people the period. The GDPR requires you to tell data subjects how long their data will be stored, or the criteria used to determine that period (Article 13(2)(a) GDPR, and Article 14(2)(a) GDPR for data that, like logs, is not collected from the data subject), and to include in the record of processing activities, where possible, the envisaged time limits for erasure (Article 30(1)(f) GDPR). A separate entry for "logs of the Work and Income AI system" with a period and its justification makes that concrete; see Article 30 GDPR for AI.
On the legal basis: for retention during the minimum period, Article 6(1)(c) GDPR (legal obligation) is on our reading the obvious basis, because Article 26(6) AI Act lays down a retention obligation in Union law (Article 6(3) GDPR). For longer retention, the municipality relies on the basis of the processing itself, usually its public task, and must justify the necessity. The AI Act expressly provides that Union law on the protection of personal data applies to personal data processed in connection with the Regulation, and that the GDPR is not affected (Article 2(7)).
Security: a log is only evidence if it is accurate
Logs are an attractive target and a fragile piece of evidence at the same time. The GDPR requires appropriate technical and organisational measures to ensure a level of security appropriate to the risk, including where appropriate pseudonymisation and encryption (Article 32(1) GDPR), and processing that ensures integrity and confidentiality (Article 5(1)(f) GDPR). Under integrity and confidentiality, the EDPB lists access control management among other elements: only authorised personnel who need it get access, with as few people and as little data per task as possible. For back-ups and logs, the same point expressly mentions pseudonymisation, protection against unauthorised and accidental access and change, regular review, and audit trails and event monitoring as a routine security control (Guidelines 4/2019, point 85).
Translated to logs: store them so that neither the system nor its users can alter them afterwards, limit who can read them to system administration, the privacy officer and whoever monitors, log who consults the logs, and pseudonymise where monitoring does not need names. A log that anyone can edit proves nothing in an objection, an incident or a question from the authority.
Logs in human oversight, monitoring and incidents
Logs are not an archiving duty for their own sake. Article 12(2) expressly links them to monitoring; on our reading they are also the evidence that human oversight and incident handling work.
Human oversight. The municipality assigns human oversight to staff with the necessary competence, training and authority (Article 26(2)). The system must enable them to decide, in any particular situation, not to use the output, to disregard it, or to override or reverse it (Article 14(4)(d)). Whether that happens in practice only shows in the logs: how often do caseworkers follow the recommendation, and when do they depart from it? A deviation rate of zero is not proof of a good system; it may also point to blind reliance.
Monitoring and suspension. The municipality monitors the operation on the basis of the instructions for use and informs the provider where relevant (Article 26(5)). If it has reason to consider that use in accordance with the instructions may result in a risk within the meaning of Article 79(1), it informs the provider or distributor and the market surveillance authority without undue delay, and suspends the use. Suspending is not optional. Article 12(2)(c) expressly names this monitoring as a purpose of logging.
Serious incidents. If the municipality identifies a serious incident, it immediately informs first the provider, and then the importer or distributor and the market surveillance authorities (Article 26(5)). The provider reports the incident to the market surveillance authority immediately after establishing a causal link or the reasonable likelihood of one, and in any event not later than 15 days after it or the deployer becomes aware of the incident; in the event of a widespread infringement or a serious and irreversible disruption of critical infrastructure not later than two days, and in the event of death not later than ten days (Article 73(2) to (4)). The provider then investigates and does not alter the system in a way that may affect any subsequent evaluation of the causes without informing the authorities first (paragraph 6). On our reading, this means for the municipality: suspend the automatic deletion of the logs relating to the incident as soon as an incident comes into view, and set out that exception to the retention period in advance in the retention policy.
Monitoring by the provider. The provider actively collects and analyses data on the performance of the system throughout its lifetime, including data that deployers may provide (Article 72(2)). Agree which log data you share, in what form and whether it can be aggregated. If that data contains personal data, that is a disclosure that needs a legal basis.
Logs and the right to explanation
Any affected person subject to a decision taken by the deployer on the basis of the output of a high-risk AI system listed in Annex III (with the exception of point 2), which produces legal effects or similarly significantly affects them in a way that they consider to have an adverse impact on their health, safety or fundamental rights, has the right to obtain a clear and meaningful explanation of the role of the AI system in the decision-making procedure and the main elements of the decision (Article 86(1) AI Act). The right does not apply to the use of AI systems for which exceptions from, or restrictions to, that obligation follow from Union or national law (paragraph 2). That right applies only to the extent that Union law does not already provide for it (paragraph 3). On our reading, that may be the case for solely automated decisions under Article 22 GDPR; for a recommendation that a caseworker reviews, it often is not. The municipality must also inform applicants that the system is used on them (Article 26(11)).
Without logs, the municipality cannot give that explanation. Which model version, which recommendation, and what did the caseworker do with it? That is in the log of that one decision. This is a strong argument for keeping the logs of a decision for as long as an explanation can still be requested or an objection can be pending, and that argument belongs in the justification of the period.
From when?
Articles 12, 19 and 26 sit in Chapter III, Sections 2 and 3. For high-risk systems under Annex III they apply from 2 December 2027, for systems under Annex I from 2 August 2028 (Article 113, third paragraph, point (c), as amended by Regulation (EU) 2026/1744). A system for assessing applications for social assistance falls, on our reading, under Annex III, point 5(a): AI used by or on behalf of public authorities to evaluate the eligibility of natural persons for essential public assistance benefits, or to grant, reduce, revoke or reclaim such benefits. If such a system performs profiling of natural persons, it is always considered high-risk (Article 6(3), last subparagraph).
If the system (that type and model) was placed on the market or put into service before 2 December 2027, the high-risk obligations only apply if it is subject to significant changes in its design from that date. Providers and deployers of systems intended to be used by public authorities must in any event comply with the Regulation by 2 August 2030 (Article 111(2)). A municipality therefore has time, but no reason to wait: the GDPR already applies to the logs the system creates today, and logging that has to be rebuilt later costs more than logging set up properly from the start.
Example 2: the bank and credit assessment
A bank that uses AI to evaluate the creditworthiness of individuals or establish their credit score uses a high-risk system under Annex III, point 5(b) (with the exception of AI used to detect financial fraud). Two particular rules then apply to the logs.
The logs go into existing documentation. A deployer that is a financial institution subject to requirements regarding its internal governance, arrangements or processes under Union financial services law keeps the logs as part of the documentation it already keeps under that law (Article 26(6), second subparagraph). For a bank that develops the system itself and is therefore the provider, the same rule is in Article 19(2). The Regulation does not say in so many words whether the six-month minimum continues to apply alongside. The safe line is to follow both: at least six months and the periods under financial services law.
Monitoring runs through governance. For the same financial institutions, the monitoring obligation of Article 26(5) is deemed fulfilled by complying with the internal governance rules under financial services law (Article 26(5), second subparagraph). On our reading, that does not remove the duty to inform the provider and the authority in case of a risk or a serious incident; the safe line is to set up that reporting route separately.
The GDPR applies to the bank just as it does to the municipality. Pankki S was itself about a bank: a customer may know when and why their data was consulted, even if that has to come from log files. In addition, a bank deploying a point 5(b) system carries out a fundamental rights impact assessment (Article 27(1)); see the FRIA template and, for the credit decision itself, from scoring algorithm to transparent credit decision.
What to do now
- Map the logging. Ask the vendor which events the system logs, where the logs are stored, who has access and how long it keeps them. Test that against the three purposes of Article 12(2).
- Bring the logs under your control. Set out access, export, retention and deletion in the contract and the data processing agreement.
- Set a period per type of log. At least six months, longer only with a justification: objection and explanation periods, monitoring, other law. Automate deletion, and set out when deletion is suspended, for example during an incident or a pending objection.
- Secure the logs. Store them so they cannot be altered, restrict access, record consultations of the logs themselves, pseudonymise where possible.
- Update the record, privacy notice and DPIA. Include the logs in the record of processing activities with their period, inform data subjects about the period, and use the information from the instructions for use for the DPIA (Article 26(9) AI Act and Article 35 GDPR).
- Prepare for access and explanation. Make sure you can retrieve the logs per person for an access request and for an explanation request under Article 86.
How the GDPR and the AI Act interact more broadly is set out on the page on the GDPR alongside the AI Act. Unsure whether your system is high-risk and which obligations then apply? With the case check you describe the application in a few questions and see which AI Act and GDPR obligations come into view.
Frequently asked questions about logs of high-risk AI
Sources
Newsletter
Every Tuesday, the AI Act week ahead in 5 minutes
A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.
Practical and short · No spam · One-click unsubscribe