Predictive policing: Article 5(1)(d) and crime risk assessment based solely on profiling

What practice does point (d) describe?

Point (d) describes AI systems that assess or predict the risk of a person committing a criminal offence. The assessment can take the form of a score, a category, a ranking or a flag, and it can be presented as a prediction, a risk level or a recommendation for attention. The prohibited basis is sole reliance on profiling or on personality trait and characteristic assessment. Profiling, as defined in the Act by reference to data protection law, means automated processing of personal data to evaluate personal aspects, including behaviour, characteristics and movements. The record should describe the inputs, the profiling involved, the output, and the decision or action the output informs, because the prohibition is about the assessment the system performs, not about the label on the tool.

What does based solely on mean in practice?

Based solely on means the profiling or personality assessment is the only basis for the risk assessment. A system that combines profiling with objective and verifiable facts directly linked to a criminal activity, in a human assessment that the system supports, follows a different pattern. The distinction is about the actual basis of the assessment, not about whether other data exist somewhere in the organisation. If the human decision-maker in practice relies only on the score, or cannot access or evaluate the objective facts, the assessment may still be solely based on profiling in substance. Workflow evidence, including what the human sees, what they can change, and what they actually do, is therefore as important as the system design.

For the broader regulatory framework, see prohibited practices under Article 5.

What is the human assessment exception?

The exception covers AI systems used to support the human assessment of the involvement of a person in a criminal activity, where that human assessment is based on objective and verifiable facts directly linked to a criminal activity. Three elements must coincide: a human assessment, objective and verifiable facts, and a direct link to a criminal activity. Facts such as a specific incident report, verified evidence, or a documented link to an ongoing investigation illustrate the kind of material the exception describes. General characteristics, past behaviour unrelated to a specific activity, or inferences about personality do not meet the test on their own. The exception should be evidenced with the facts relied on and the human's reasoning, not asserted from the presence of a human in the workflow.

How does point (d) relate to Annex III law enforcement uses?

Annex III lists law enforcement use cases that can be high-risk, including individual risk assessment of offending, polygraphs and similar tools, evidence evaluation, and crime analytics for patterns. A system that falls outside point (d) because it supports a human assessment on objective facts can still be high-risk under Annex III, with the full Chapter III regime to follow. A system that falls inside point (d) is prohibited and should not be treated as a high-risk system to be managed. The two analyses must be done in order: point (d) first, Annex III second. Conflating them risks either prohibiting a lawful support tool or, more seriously, managing a prohibited prediction as if it were a compliant high-risk system.

What should the screening record contain?

The record should name the criminal risk being assessed, the inputs and their sources, the profiling involved, the personality traits or characteristics assessed, the output, the human who receives it, the objective facts available to that human, and the decision that follows. It should describe whether the human can depart from the output, what information supports a departure, and how the decision is recorded. Where the exception is relied on, the record should cite the specific objective facts and show the direct link to a criminal activity. Where no such facts exist, the record should state that plainly, because the absence of objective facts is itself the finding that keeps the system inside the prohibition.

Continue with social scoring prohibition.

What does this guide not determine?

This guide explains the prohibition and its exception in outline. It does not decide whether any particular law enforcement, security or risk tool is prohibited, whether its human assessment meets the exception, or whether its facts are objective and directly linked to a criminal activity. Those questions need the real inputs, the real workflow, and the real decisions, read against the consolidated text and the applicable law enforcement and data protection rules. Tools that assess the risk of offending should be confirmed with a competent authority or legal adviser before deployment, given the seriousness of a wrong conclusion in either direction.

Frequently asked questions

Continue through this AI Act cluster

• prohibited practices under Article 5
• social scoring prohibition
• Annex III law enforcement, migration and justice uses
• the risk-based approach
• high-risk AI systems
• Understand AI Act risk categories