DPIA for AI applications

According to the Autoriteit Persoonsgegevens (AP), an organisation must first carry out a DPIA if processing personal data is likely to pose a high privacy risk. On algorithms, the AP says that if you want to use algorithmic systems that process personal data, a DPIA is mandatory. This also applies to pilots, tests and trial projects.

The AP refers to Article 35 of the GDPR. A DPIA is part of your broader data governance.

The AP identifies three sources of guidance: the GDPR itself, the AP’s DPIA list and nine criteria set by EU privacy regulators. As a rule of thumb, you should carry out a DPIA if your processing meets two or more criteria. Examples include automated decisions, profiling, sensitive data, linked databases and new technology.

If you meet none or only one of the criteria, the AP says you must be able to explain clearly why you are not carrying out a DPIA.

According to the AP, it must include at least a systematic description of the processing and its purposes, an assessment of necessity and proportionality, an assessment of privacy risks and the proposed measures.

For algorithms, the AP calls for particular attention to technical aspects such as the choice of algorithm and bias, to transparency and explainability, and to rights such as erasure and rectification. The Model DPIA Rijksdienst sets out a DPIA in seventeen points.

The data controller must ensure that a DPIA is carried out and remains ultimately responsible, even if someone else carries it out. If your organisation has a data protection officer (DPO), the AP says you must seek their advice and include it in the report.

A processor must support you by providing the information you need. That is where our work fits in: providing the technical description of the data, model and measures.

According to the AP, the FRIA, the fundamental rights impact assessment under the AI Act, complements the DPIA. A DPIA focuses on privacy risks; a FRIA looks more broadly at all fundamental rights when a high-risk AI system is deployed. From December 2027, a FRIA will be mandatory for, among others, public-sector organisations planning to use certain high-risk systems.

Your legal adviser can assess what applies to you; see also the AI Act.

The AP describes a DPIA not as a one-off task but as a continuous process. Changes to the processing, the risks or the context require a new DPIA. In AI, this could mean a new model, new training data or a different use of the output.

That is why we link model version control and monitoring to the DPIA through MLOps. We also work with LUMC on monitoring and maintenance in this way.

The AP says you must then consult the AP before starting the processing. This is called prior consultation. For each high residual risk, describe the circumstances in which it arises, how likely it remains despite your measures and what harm people could suffer.

You can read about who makes which AI decisions under AI governance. We describe an AI management system under ISO 42001.

If the AI application does not process personal data, or if you have purchased software and the supplier provides the technical information your DPO requests. In those cases, your organisation’s existing DPIA process is sufficient.

If the algorithms must also be described publicly, read our page on the Algorithm Register.