How do you run a DPIA for an AI system?

Cover the four elements GDPR Article 35 requires: describe the processing, assess necessity and proportionality, assess the risks to people, and set out the measures. Then add the AI-specific layer: where personal data enters the model, what the model can leak or infer, whether any anonymity claim holds, and how you handle hallucinated personal data.

Under GDPR Article 35, a data protection impact assessment is required where processing is likely to result in a high risk to people's rights and freedoms. Many AI systems that process personal data meet that threshold. Article 35(7) sets the minimum content: a systematic description of the processing and its purposes, an assessment of necessity and proportionality, an assessment of the risks to data subjects, and the measures planned to address them.

For an AI system, each element needs an AI-specific layer:

  1. Describe the system and its data flows. Cover training and fine-tuning data, retrieval sources, prompts, outputs and logs, not only the production database.
  2. Test necessity and proportionality. Ask whether you need personal data to train or ground the model, and whether pseudonymised or synthetic data would do.
  3. Assess model-level risks. These include memorisation and training-data extraction, membership inference, model inversion, poisoning, insider access to logs and embeddings, and hallucinated personal data. ISO/IEC 27091 (publication estimated for February 2027) gives a structured method for this analysis.
  4. Check any anonymity claim. EDPB Opinion 28/2024 says AI models trained on personal data cannot always be considered anonymous, and that anonymity is assessed case by case. Document the extraction and inference tests behind your conclusion.
  5. Plan the measures and the residual risk. Examples are access limits, pseudonymisation before training, output filtering, grounding, retention limits and a process for rectification and erasure requests about model outputs. If high residual risk remains, Article 36 requires prior consultation with the supervisory authority.

Review the DPIA whenever the model, the training data or the use case changes.

Read more: ISO/IEC 27091 guide · How ISO/IEC 27701 supports GDPR compliance.

Browse all FAQs →

Full knowledge base