Predictive Policing AI Under the EU AI Act: Where the Prohibited Line Sits
Predictive policing AI that predicts offending solely on profiling is prohibited (Art 5(1)(d)); risk tools supporting human judgment are high-risk.
Predictive policing is the use case where the EU AI Act draws its sharpest line — and where getting the classification wrong is most costly. Some predictive-policing AI is prohibited outright under Article 5; some is high-risk under Annex III. The difference turns on whether the system predicts criminal behaviour solely from profiling, or whether it supports a human assessment grounded in objective, verifiable facts. This page sets out the boundary precisely.
The prohibited core: predicting offending from profiling (Article 5(1)(d))
Article 5(1)(d) prohibits AI used to make risk assessments of natural persons in order to assess or predict the likelihood of a person committing a criminal offence, based solely on profiling or on assessing personality traits and characteristics. This is the classic "predictive policing" the Act bans: software that flags an individual as a future offender from their profile, history pattern, or characteristics, without a human judgment tied to objective facts. It has been prohibited since 2 February 2025, and breaches sit in the top penalty tier — up to €35 million or 7% of total worldwide annual turnover (Article 99(3)).
There is a narrow exception: the prohibition does not cover AI used to support a human assessment that is already based on objective and verifiable facts directly linked to a criminal activity. That exception is the bridge to the high-risk category.
The high-risk side: law-enforcement risk tools (Annex III point 6)
Where a system supports — rather than replaces — a human assessment grounded in objective, verifiable facts directly linked to criminal activity, it is not prohibited. Instead it falls under Annex III point 6 (law enforcement), which lists AI intended to be used by or on behalf of law-enforcement authorities to assess the risk of a person offending or re-offending, or as a polygraph, or to evaluate the reliability of evidence, among others. These are high-risk uses, carrying the full provider/deployer regime — not a ban.
The test, in one sentence: prediction of offending from profiling alone is prohibited; risk assessment that informs a human decision grounded in objective facts is high-risk.
A related prohibition: real-time public biometric ID
Predictive-policing programmes often pair with biometric surveillance, which has its own rule. Real-time remote biometric identification in publicly accessible spaces for law-enforcement purposes is prohibited under Article 5(1)(h), save for narrow, authorised exceptions. Untargeted scraping of facial images to build recognition databases (5(1)(e)) and biometric categorisation inferring sensitive attributes (5(1)(g)) are likewise prohibited. Post (non-real-time) biometric identification, by contrast, is Annex III point 1 high-risk. Keep these distinct from the predictive-policing analysis.
The FRIA and the public-body context
Because predictive-policing and law-enforcement risk tools are typically deployed by public authorities, the Article 27 Fundamental Rights Impact Assessment is mandatory: public-body and public-service deployers must assess the impact on fundamental rights before first use. For a high-risk law-enforcement tool this is not optional best practice — it is a required step alongside the Article 26 deployer duties.
Worked example: a regional police authority
A regional police authority evaluates two tools. Tool A scores individuals on their likelihood to offend from demographic and behavioural profiling, with officers acting on the score. That is prohibited under Article 5(1)(d) — there is no compliance route; it cannot be deployed. Tool B helps analysts assess re-offending risk for individuals already charged, drawing on objective case facts and feeding a documented human decision. That is Annex III point 6 high-risk: lawful, but it requires the provider stack, human oversight under Article 14, and a Fundamental Rights Impact Assessment under Article 27 before use. Same domain, opposite outcomes — the line is profiling-alone versus fact-grounded human judgment.
How Confir helps
Confir's classification logic draws exactly this boundary: it tests a law-enforcement or risk-scoring use against the Article 5(1)(d) prohibition first, and only then against Annex III point 6, so a banned profiling tool is flagged as prohibited rather than mistakenly scoped as "high-risk to be documented." For genuinely high-risk uses it assembles the provider stack and the Article 27 FRIA. The engine is deterministic and rule-based, with each finding cited to the Article and Annex point that fired.
Frequently asked questions
Is predictive policing AI banned under the EU AI Act?
Partly. AI that assesses or predicts the likelihood of a person committing a criminal offence based solely on profiling or personality traits is prohibited under Article 5(1)(d), in force since 2 February 2025. AI that supports a human assessment grounded in objective, verifiable facts directly linked to criminal activity is not prohibited — it is high-risk under Annex III point 6.
What is the difference between prohibited and high-risk here?
The line is profiling versus fact-grounded human judgment. Predicting offending from profiling alone is prohibited (Article 5(1)(d)). A risk tool that informs a human decision based on objective, verifiable facts is high-risk under Annex III point 6, carrying the provider/deployer obligations rather than a ban.
Does the prohibition cover biometric surveillance too?
Those are separate prohibitions. Real-time remote biometric identification in public spaces for law enforcement is prohibited under Article 5(1)(h) (with narrow exceptions); untargeted facial scraping (5(1)(e)) and sensitive biometric categorisation (5(1)(g)) are also prohibited. Post, non-real-time biometric identification is Annex III point 1 high-risk.
Is a FRIA required for law-enforcement risk tools?
Yes, for the high-risk ones. Because they are deployed by public authorities, the Article 27 Fundamental Rights Impact Assessment is mandatory before first use, alongside the Article 26 deployer duties and Article 14 human oversight.
What are the penalties?
A prohibited predictive-policing use falls in the top tier — up to €35 million or 7% of total worldwide annual turnover under Article 99(3). Breaches of high-risk obligations fall under Article 99(4), up to €15 million or 3%. SMEs and start-ups are capped at the lower figure under Article 99(6).
Related guides
- Article 5 prohibited practices
- Annex III high-risk use cases
- The FRIA under Article 27
- How AI risk classification works
- How Article 6 classifies high-risk AI
Manage your EU AI Act compliance in one place
Confir automates risk classification, technical documentation, and audit trails for any company. No consultants. No 6-month projects. 14-day free trial.
Start free trial →