AI screening and GDPR when a legal basis for AI is not the same as lawful use
Bill L 99 gives Danish public authorities a legal basis for AI with decision support and profiling, but not for automated decisions. Datatilsynet's consultation response sets out what has to be settled before AI screening can start.
On 11 November 2026, bill L 99 was introduced in the Danish parliament. Its title translates as the Act on public authorities' and bodies' development and use of AI systems when processing personal data, and it is commonly known as the AI legal basis act. Two days later, on 13 November, the Danish Data Protection Agency, Datatilsynet, published a news item that brings together its two consultation responses. The message is hard to misread. A legal basis on its own does not make the processing of personal data lawful. Public employers considering AI for candidate screening would do well to read what the regulator actually says.
What the bill does and does not offer
According to Datatilsynet's news item, the bill gives public authorities and bodies a general legal basis under data protection law to develop and use AI systems that process personal data. That includes decision support and profiling, meaning the evaluation of certain personal aspects of the individual concerned. What the bill does not provide is a basis for automated individual decisions under Article 22 of the GDPR. Even if the bill passes as introduced, a public employer would not be allowed to let a system reject applicants on its own.
Four things that have to be settled first
Datatilsynet filed its first consultation response on 3 March 2026 and a second one on 9 October 2026, after the draft had been revised. In the later response the agency repeats that a legal basis is only one of several conditions a controller has to assess before processing personal data. Four themes come back throughout. Security of processing, the rights of data subjects, transfers of personal data to third countries, and the risk of purpose creep. In the agency's wording, these requirements must be clarified before the processing begins, not along the way.
What the response says about decision support
The part that speaks most directly to recruitment is the section on decision support. The revised draft states that human involvement has to be more than a formality, and the explanatory notes list concrete measures. It must be clear who holds decision-making authority, deviations from the system's recommendations must be documented, and use of the system must be logged. Datatilsynet welcomes this, then adds three warnings.
First, the person who makes the decision must be able to evaluate its factual basis, which can be difficult with complex and opaque systems. Second, the controller must examine and address bias, which may originate in the training data of the underlying model and can lead to discrimination. Third, the agency warns about automation bias, where a caseworker gives the system's assessment more weight than their own judgement. When that happens, the system is in practice deciding, and the decision is automated whatever the written procedure says.
The agency also notes that profiling used for decision support is in itself considered intrusive, even where no automated decision is taken.
Development is a purpose of its own
A second point that is easy to miss during procurement is the distinction between developing a system and using it. Datatilsynet maintains that developing or further developing an AI system is a separate purpose under the GDPR, even when the long-term goal is operational use. The controller therefore has to state the purpose for each phase of the system's life cycle and include each phase in the data protection impact assessment. According to the agency, such an assessment will almost always be required before an AI system that processes personal data is developed or operated.
The response also points out that standardised AI products often leave the customer little room to adjust contracts or functionality. The advice is to find out before procurement whether the necessary changes are possible at all, and to budget time and money for alternative designs. Nor does the data protection work end at go-live, because AI services and their contracts change often.
What this means for candidate screening
The consultation response is not about recruitment, and neither the bill nor Datatilsynet mentions HR or hiring. Whether a particular recruitment process at a public employer falls within the bill is for the authority itself to assess. But the questions the agency puts to decision support are the same questions a recruitment team should put to a screening tool. Who holds the decision-making authority when the system ranks candidates? Can the person responsible explain why a candidate sits where she sits? Is there documentation of how often the team departs from the system's recommendation? And is the profiling of applicants described as a distinct intervention in the impact assessment?
Datatilsynet is now waiting for the political process in parliament and encourages all controllers to read the response. Several of the requirements it describes follow, in the agency's view, from the GDPR itself rather than from the bill.
Source: https://www.datatilsynet.dk/presse-og-nyheder/nyhedsarkiv/2026/nov/datatilsynets-hoeringssvar-til-ai-hjemmelsloven