Data protection impact assessment for AI

What it does

It identifies the privacy risks of a system while they can still be designed out. The assessment describes the processing, tests whether it is necessary and proportionate, identifies the risks to the people whose data is used, and records the measures to reduce them and the risk that remains. The business owner of the processing is accountable; the data protection officer advises and reviews. It is done at design, signed off before processing starts and revisited on change.

How to implement it

• Screen every AI proposal at intake for the factors that require an assessment.
• Describe the data flows fully, including training, prompts, logs, providers and transfers.
• Address the AI-specific questions: inference of sensitive attributes, memorisation, accuracy of outputs, automated decisions and the ability to honour rights.
• Seek the views of affected people or their representatives where practicable.
• Record measures with owners, and escalate high residual risk to the regulator where the law requires.
• Review on any material change and at a set interval.

Evidence it operates

Design: screening criteria and an assessment template with AI-specific sections. Operation: an assessment for each qualifying system, dated before go-live, with the data protection officer's advice; measures tracked to completion; a later version reflecting a change to the system.

Where it fits

At design, alongside the wider AI impact assessment process and triggered by AI use-case intake and approval. It is the primary control for missing DPIA for high-risk AI processing and, with lawful basis review, for inference of sensitive attributes.

Limits

An assessment written to justify a decision already made finds little. It depends on an accurate description of the system, which suppliers do not always provide. It considers data protection harms and not every harm a system can cause, which is why it sits beside the wider impact assessment and does not replace it.