Denmark's data protection authority told a government agency that it could train an artificial intelligence system on roughly 3,000 historical applications for a student disability supplement, but that a notice on its website would not be enough to tell those applicants what was happening to their files. The opinion, dated June 12, 2024 and filed under journal number 2024-212-0371, came from the Danish Data Protection Agency (Datatilsynet) in response to a request from the Danish Agency for Higher Education and Science. It drew a sharp line between building the tool and running it: existing student aid law covered development, the regulator found, while operating the system would need "a relatively clear supplementary national legal basis." Twelve months later the Danish parliament supplied one.

In Short

A Danish government office wanted to teach a computer program to check whether students applying for extra disability money had sent in enough paperwork, using about 3,000 old applications to train it. Denmark's privacy regulator said the office was allowed to build the program, but it had to write to each of those people directly instead of just posting a notice on its website, and it would need a clearer law before using the tool on new applications. Denmark passed that law in June 2025, which shows how a privacy regulator's view of AI training can end up shaping both what organisations tell people and what lawmakers write down.

What the agency wanted to build

The Danish Agency for Higher Education and Science runs SU, Denmark's state educational support scheme of student grants and loans. One element of SU is a supplement for students in higher education or vocational training who, because of a permanent physical or mental disability, have very significant limitations in their ability to take up paid work. Section 7(3) of the SU Act creates the supplement, and section 35 of the SU executive order sets out how the agency assesses it: first the nature and degree of the disability, then what that disability means for the applicant's ability to work.

According to the opinion, the agency planned to develop the system with cBrain, a Danish software company. The purpose was narrow. The tool would check whether an incoming application for the disability supplement was accompanied by sufficient documentation, judging both the amount and the type of material attached. If the system concluded that the documentation fell short, the case would be placed for manual screening and the applicant would be asked for more.

The agency stressed that the tool would not act as decision support. It would be used "only in connection with the initial screening of the enclosed documentation," according to the opinion. Whether an applicant ultimately receives the supplement would remain a separate assessment.

The training data was anything but narrow. According to the opinion, the system would be trained on received and decided disability supplement applications together with their documentation, covering:

  • the applicant's name, personal identity number and contact information
  • educational information
  • income information
  • whether the applicant had previously received SU, supplementary grants or disability supplements
  • employment information
  • information about health, private and social conditions
  • other information relevant to awarding the supplement

Roughly 3,000 historical cases would form the initial dataset. Future applications would be fed into the ongoing training of the system, according to the opinion. The regulator noted that development would therefore involve special category data, health information in particular.

Translated from Danish, the regulator's characterisation of the population is blunt. Students who need a disability supplement in order to complete an education are, in its words, "a vulnerable target group."

Building the tool: existing law was enough

The agency told the regulator it would rely on Article 6(1)(e) of the GDPR, which covers processing necessary for a task carried out in the public interest or in the exercise of official authority, and on Article 9(2)(g), which permits processing of sensitive data for reasons of substantial public interest. Both provisions require what Danish law calls a supplementary national legal basis: a law that obliges or entitles the authority to carry out a specific task.

The agency pointed to section 7(3) of the SU Act and section 35(1)-(4) of the SU executive order, which oblige it to decide on supplementary grants for permanent disability. Datatilsynet accepted that argument for the development phase. "Thus, public authorities can often design, develop and test AI solutions that can support the authority in carrying out this task, within the framework of the legislation that obliges or entitles the authority to perform a specific task," the opinion states.

The deciding factor was impact. The regulator "placed particular emphasis on the fact that the development of the solution does not entail direct consequences for the citizens," according to the opinion. Training a model on old files, in other words, does not change anyone's grant.

That reasoning came with a condition attached. Authorities, the regulator wrote, should assess "the entire life cycle of the AI solution" before starting development, to make sure they have also identified a legal basis for putting the system into operation, or to decide whether legislation is needed. One option it named explicitly was for the legislator to lay down clear national rules that could serve as the additional legal basis under Article 6(2) and (3).

The opinion built on an earlier one. On November 17, 2023, Datatilsynet gave the City of Copenhagen a statement (journal number 2023-212-0015) on the legal basis for developing and operating an AI solution in the health and care sector, and the June 2024 opinion refers to that statement for the detailed requirements a supplementary legal basis must meet when personal data is used to develop and train AI.

A website notice was not enough

The most transferable part of the opinion concerns transparency. Article 13(3) of the GDPR requires a controller that intends to process personal data for a purpose other than the one for which it was collected to inform the people concerned before that further processing begins. The applicants had sent their files to obtain a supplement, not to train a screening model. Using those files for training raises a question of purpose limitation, and it triggers the notice duty.

The agency proposed to meet that duty with information on su.dk, its student support website. It gave five reasons, according to the opinion: the number of people was relatively large, at approximately 3,000; informing them individually would be resource-intensive; the processing was limited; its purpose was to improve processing times; and, in the agency's view, it would have no direct consequences for the people involved.

Datatilsynet was not persuaded. "It is the assessment of the Danish Data Protection Agency that these considerations cannot justify that the duty of disclosure is not fulfilled directly with the data subjects," the opinion states.

Its reasoning was practical rather than theoretical. The agency already held contact details for every applicant, and the regulator noted that it "has the opportunity to provide the data subjects with the information required by means of digital or physical mail."

The regulator also checked the available exemptions and found none that applied. Section 22(1) of the Danish Data Protection Act allows Article 13 to be set aside where the data subject's interest in the information is outweighed by decisive private interests. Section 23 lifts the Article 13(3) duty when public authorities further process data under rules laid down under section 5(3) of that Act. According to the opinion, the agency had not identified a legal basis that would exempt it under section 23, and Datatilsynet had not "been able to identify such a legal basis" either.

The structure of the GDPR explains why cost carried so little weight. Article 14, which governs data obtained from third parties, contains an exemption where informing people would involve disproportionate effort. Article 13, which governs data collected from the person directly, has no such exemption. The applicants had submitted their own files.

Running the tool: "a relatively clear" basis

Operation was a different matter. The agency had not, at the time of the request, made a final assessment of whether running the system would be lawful, according to the opinion. Datatilsynet nonetheless set out its view for the agency to use in that assessment.

The opinion opens that section with a concession that is easy to miss. "The Danish Data Protection Agency notes that the use of AI solutions cannot generally be said to be intrusive towards the citizen," it states. Citizen-facing public tasks are different, however, and AI used as decision support in administrative case processing will often be intrusive.

The screening tool sits in an awkward middle. Because it would only check whether an application is complete, the processing "does not have a direct impact on the applicants' ability to be awarded disability supplement," according to the opinion. Yet the volume of personal data, including sensitive data, pointed the other way.

Why AI changes the calculation

The regulator then explained why AI raises the bar for the clarity of the law that authorises it. "The use of AI means that AI solutions can learn, find correlations and carry out probability analyses and draw conclusions far beyond what a physical case worker would be able to do," the opinion states. "The use of AI in administrative case processing is thus fundamentally different from the traditional case processing that has been the norm until now."

Add the vulnerability of the applicant group, and the conclusion followed. "Against this background, it is the assessment of the Danish Data Protection Agency that a relatively clear supplementary national legal basis is required for the operation of the AI solution," the opinion states.

Impact assessment expected

The agency was also reminded of Article 35. If processing is likely to result in a high risk to the rights and freedoms of individuals, a data protection impact assessment must be carried out before it begins. "In the opinion of the Danish Data Protection Agency, this is often the case when developing and operating AI solutions," the opinion states. It pointed the agency to section 6 of Datatilsynet's guidance on public authorities' use of artificial intelligence and to the regulator's spreadsheet template for impact assessments of AI solutions.

What Denmark did next

The legislative record shows the opinion did not sit in a drawer. According to the explanatory remarks to bill L 152, a taskforce on the administration of the disability supplement recommended digital screening tools that would automatically sort incoming applications into adequately and inadequately documented ones. The bill's remarks put the scale of the problem in numbers: information was missing in 64 percent of applications processed in 2024, and the average processing time for all disability supplement applications was about 6.1 months that year, rising to 6.8 months for higher education applicants and up to about 10 months where extra documentation had to be requested. The supplement itself was worth DKK 10,078 a month for higher education students and DKK 6,321 for vocational students at 2025 levels, according to the bill.

The ministry consulted on a draft of the bill from December 2, 2024 to January 7, 2025, according to its consultation memo. Datatilsynet responded. According to the memo, the regulator said it could not assess, on the basis of the proposed enabling power alone, whether the legal basis would be sufficiently clear and precise. It asked the ministry to be particularly attentive when using the power, expected a risk assessment under Article 24(1) of the GDPR and an impact assessment under Article 35 where the risk is high before any AI solution goes into use, and expected to be consulted again when the power is used. The ministry replied that the agency and Datatilsynet had already been in dialogue about developing an AI solution before the consultation and that the regulator would be involved when the rules were drafted.

Christina Egelund, then minister for higher education and science, introduced L 152 on February 26, 2025. Its explanatory remarks cite the June 12, 2024 opinion and its call for a relatively clear supplementary legal basis. They also state that the enabling power is not intended to be used for profiling citizens or for automated decisions.

On March 20, 2025, the parliament's education and research committee asked, at the request of Heidi Bank of Venstre, how long the risk assessment and impact assessment proposed by Datatilsynet would take. Egelund's answer of March 27, 2025 gave no estimate. According to the answer, the agency was preparing a pilot project, and a decision on putting the solution into operation was expected during 2025.

Parliament adopted the bill as Act no. 674 of June 17, 2025, in force from July 1, 2025. It inserted a new section 39 a into the SU Act, under which the minister may lay down rules on the use of artificial intelligence or similar digital solutions to support case processing under the Act, including the processing of personal data that this involves. The same law allows more detailed documentation requirements for disability supplement applications and the rejection of applications where documentation is not submitted.

The rules followed within days. After a consultation launched on May 16, 2025 with a June 11 deadline, the agency signed executive order no. 813 on June 21, 2025, also in force from July 1, 2025. Its scope is wider than the disability screen. According to the order, the agency may use AI or similar digital solutions to process personal data from SU applications where this is necessary to support its case processing, with the stated purpose of optimising case processing. The order covers the development and operation of AI for the allocation, calculation and repayment of SU, lists nine categories of personal data that may be processed, including personal identity numbers, health data and information on criminal matters, and names the sources: the agency's own registers and case systems, registers under section 39(1) of the SU Act, and publicly available sources. Data may be kept for as long as necessary for those purposes. The agency must publish on su.dk when it uses AI. PPC Land did not find an explicit prohibition on automated decisions in the order's ten sections; that limit appears in the bill's explanatory remarks.

Whether the screening tool is now running is not clear from public sources. A written question from Karina Adsbøl of the Denmark Democrats, listed by the Danish outlet Altinget on August 24, 2026, asked whether it was reasonable that students currently wait 6.2 months for an answer to a disability supplement application. The answer is not publicly available. PPC Land found no public statement confirming that the screening tool has entered operation, nor any record of whether the roughly 3,000 historical applicants were written to individually.

A summary that cites a later rule

The opinion contains an internal inconsistency. Its body, in section 5, says the agency had not yet assessed the legality of operation and that a relatively clear supplementary national legal basis was required. Its summary, published on the same page, says something else: processing for operation "can, according to the Data Protection Agency's assessment, take place with reference to existing provisions in the SU Act, the SU Order and the AI-SU Order."

The summary goes further. "In particular, the Danish Data Protection Agency has attached importance to the fact that the AI-SU Executive Order clearly stipulates that the processing of personal data may take place in connection with applications or other processing of data relating to SU when it is necessary to support the Agency's case processing," it states.

The only executive order matching that description is no. 813, signed on June 21, 2025, more than a year after the opinion's date. The bill introduced in February 2025 quotes the opinion's demand for a clearer legal basis, which indicates that the body text dates from 2024. The page shows only the June 12, 2024 date and carries no note of a later revision. Whether Datatilsynet updated the summary after the order took effect, or issued a further assessment of the operating phase that is not published separately, cannot be established from the page itself.

Why this matters beyond Danish student aid

The legal basis at the centre of the opinion, Article 6(1)(e), is available only to public authorities and bodies exercising official authority. Advertisers, platforms and ad tech companies rely instead on consent or legitimate interest. Much of the reasoning travels nonetheless.

The transparency finding applies to every controller. Any organisation that collected data directly from people for one purpose and later decides to train a model on it faces the same Article 13(3) duty, and the same absence of a disproportionate-effort exemption. Datatilsynet's position was that a population of 3,000, the cost of individual contact and the aim of faster processing did not justify a website notice when the controller already held contact details. Retailers, subscription businesses and publishers that hold customer email addresses are in a comparable position, whatever lawful basis they use.

The industry's most visible test case took a different shape. Meta announced on April 7, 2025 that it would use public posts and comments from adult users in the EU to train its AI models, giving users until May 27, 2025 to object. Meta relied on legitimate interest rather than a public task, and its training data was public content rather than benefit applications, but the underlying question of how people learn that their data is being repurposed is the same one the Danish regulator answered in June 2024.

The split between development and operation also echoes European guidance. The European Data Protection Board's opinion on AI models in December 2024 treated development and deployment as distinct phases and said that unlawfully processed training data can affect the later use of a model. A review of 19 regulatory guidelines covered by PPC Land in March 2026 found legitimate interest had become the dominant basis for AI training, while regulators still diverged on how to apply it.

Sensitive data in training sets remains unsettled at EU level. Drafts of the Commission's Digital Omnibus reported in November 2025 would allow AI companies to process special categories of data, including health information, for training and operating AI. The EDPB and the European Data Protection Supervisor rejected parts of that plan in February 2026, while welcoming an exemption for incidental and residual sensitive data that is conditioned on deletion being impossible or disproportionate. A Council compromise dated September 3, 2026 added a new Article 9(2)(k) on unintentional and residual sensitive data in AI development. The Danish screening tool was never in that territory: it was designed to process health data deliberately, and it did so on the strength of a specific national law.

Impact assessments are another point of convergence. Datatilsynet said in 2024 that they are often required when developing and operating AI. The EDPB adopted its first DPIA template on March 10, 2026, with consultation open until June 9, 2026.

The AI regulation adds a further layer. Denmark was one of the first member states to legislate for the EU AI Act, with a law passed on May 8, 2025 that took effect on August 2, 2025. Annex III of the AI Act lists AI systems used by public authorities to evaluate eligibility for essential public assistance benefits and services as high-risk, while Article 6(3) excludes systems that perform narrow procedural or preparatory tasks. The June 2024 opinion does not address the AI Act at all. Under the provisional agreement of May 7, 2026, obligations for Annex III systems apply from December 2, 2027.

What the Danish case shows most clearly is sequence. A regulator assessed an AI project in mid-2024, told the authority what it could and could not rely on, and within a year the government had written a specific legal basis into statute and an executive order. Few of the commercial AI training programmes that PPC Land has covered have followed that path, because private companies cannot legislate their own legal basis. They are left with consent, legitimate interest and the transparency duties that Datatilsynet applied without exception.

Timeline

Summary

Who: The Danish Data Protection Agency (Datatilsynet) and the Danish Agency for Higher Education and Science, which planned to build the AI tool with the software company cBrain. About 3,000 past applicants for the SU disability supplement were in the training data.

What: Datatilsynet found that developing an AI tool to check whether disability supplement applications are fully documented could rest on existing student aid law, but that the agency had to inform the 3,000 applicants directly rather than through su.dk, and that operating the tool required a relatively clear national legal basis. Denmark then amended the SU Act and issued an executive order on AI in SU case processing.

When: The opinion is dated June 12, 2024, following a request of January 11, 2024. The SU Act amendment and executive order no. 813 took effect on July 1, 2025. The page's summary cites that 2025 order despite its 2024 date.

Where: Denmark, under the GDPR, the Danish Data Protection Act, the SU Act and the SU executive orders.

Why: The agency wanted to cut processing times for an application type where information was missing in 64 percent of cases in 2024. The opinion matters beyond Denmark because its transparency reasoning on reusing directly collected data for AI training applies to any controller, including advertisers and platforms that train models on customer data.