Recidivism-Prediction AI Under the EU AI Act: High-Risk or Prohibited?
AI assessing re-offending risk for law enforcement is high-risk (Annex III pt 6). Predicting offending solely from profiling is prohibited (Art 5(1)(d)).
Both are possible, and the line between them is the whole point. AI used by a law-enforcement or judicial authority to assess a known individual's risk of re-offending — built on facts about a real case and offence history — is high-risk under Annex III point 6 of the EU AI Act, Regulation (EU) 2024/1689. But AI that predicts whether a person will commit a crime solely on the basis of profiling or assessing their personality traits and characteristics is prohibited under Article 5(1)(d) — a banned practice, not a documentable high-risk one. Get the boundary wrong and you move from a compliance project to a top-tier infringement.
If you are a prison service, probation authority, police force, prosecutor or court considering a risk-assessment tool, this page draws that line and sets out the high-risk obligations — including the FRIA and the restricted-database registration the public-sector version carries.
The boundary: assessing re-offending risk vs predicting offending from profiling
The two provisions look adjacent but do different work.
Annex III point 6 designates as high-risk AI intended for use by, or on behalf of, law-enforcement authorities (or EU bodies supporting them) to assess the risk of a natural person offending or re-offending — not solely on profiling under Article 5(1)(d) — and other listed law-enforcement uses such as assessing the reliability of evidence. A recidivism tool that scores a specific, identified person against objective, verifiable facts — their offence history, the circumstances of the case, established risk factors — and feeds a human decision on bail, parole, sentencing input or supervision level, sits here. It is high-risk and lawful if the obligations are met.
Article 5(1)(d) prohibits AI used to make risk assessments of natural persons to assess or predict the risk of committing a criminal offence, based solely on profiling of a person or on assessing their personality traits and characteristics. The Act's own carve-out: the prohibition does not apply where the AI merely supports a human assessment that is itself based on objective and verifiable facts directly linked to a criminal activity. So the trigger for the ban is solely-profiling prediction of future offending detached from concrete facts — the "pre-crime" use. That is banned outright, in force since 2 February 2025, with the top penalty.
| What the tool does | Classification | Legal basis |
|---|---|---|
| Scores a known offender's re-offending risk from case facts and history, supporting a human decision | High-risk | Annex III point 6 |
| Assesses reliability of evidence for law enforcement | High-risk | Annex III point 6 |
| Predicts whether a person will offend solely from profiling or personality traits | Prohibited | Article 5(1)(d) |
| "Pre-crime" individual prediction not anchored to objective, verifiable facts | Prohibited | Article 5(1)(d) |
The practical test: is the assessment anchored to objective, verifiable facts directly linked to a criminal activity, with a human in the loop — or is it solely a profile-driven prediction of future offending? The first is high-risk; the second is prohibited. Document which side a tool sits on before you procure it.
What high-risk recidivism AI owes — FRIA and restricted-database registration
Where a tool is high-risk under point 6, the deployer — typically a public authority — carries the heavier end of the regime.
The Article 26 deployer duties apply: operate per the provider's instructions, assign competent human oversight (Article 14) by people with the authority and standing to override the score, monitor operation, and keep the logs. A risk score must never be the decision; it informs a human who decides and can disregard it.
Because the deployer is a public authority (or a body acting on its behalf), an Article 27 Fundamental Rights Impact Assessment (FRIA) is mandatory before first use — distinct from a GDPR Article 35 DPIA, though Article 27(4) lets the FRIA build on an existing DPIA.
Registration is the trap people miss. High-risk systems register in the EU database under Article 49 — but for law-enforcement, migration, asylum and border-control uses (Annex III points 1, 6 and 7), registration goes into a restricted, non-public section of the database, accessible to the Commission and national authorities rather than the public. So you still register; you register into the restricted part.
The provider side (whoever builds the tool, Article 16) owes the full Article 9–15 stack, Annex VI conformity under Article 43, the Declaration of Conformity (Article 47) and CE marking (Article 48).
Deadline. High-risk Annex III obligations were originally timetabled for 2 August 2026; the Digital Omnibus — Regulation (EU) 2026/1744, in force since 27 July 2026 — moved that to a fixed 2 December 2027. The Article 5(1)(d) prohibition, by contrast, is already in force (since 2 February 2025) and is not affected by the deferral. Plan high-risk work against 2 December 2027; treat the prohibition as live now.
Worked example: a national probation service
The Justizvollzug agency of a German Land — a public probation and prison authority — evaluates two tools for parole and supervision decisions.
- Tool A scores an identified prisoner's re-offending risk using verified offence history, the facts of the conviction, and validated risk factors, producing a band that a parole officer weighs alongside interviews and the case file. This is high-risk under Annex III point 6 and lawful if built right. The agency, as a public-body deployer, must complete an Article 27 FRIA before go-live, assign Article 14 human oversight by officers empowered to override the band, and ensure the tool is registered in the restricted section of the EU database under Article 49.
- Tool B, pitched by a different vendor, claims to flag "individuals likely to offend" across the general population from personality profiling and behavioural traits, untethered to any specific offence. That is prohibited under Article 5(1)(d) — the agency must reject it; deploying it would expose it to up to €35 million or 7% of total worldwide annual turnover (Article 99(3)).
The agency proceeds with Tool A under the full high-risk programme and refuses Tool B. As a public authority it cannot rely on any SME cap.
How Confir helps
Confir registers each law-enforcement AI use, derives the role, and tests the use against Article 5 before Annex III — so a solely-profiling "pre-crime" predictor is flagged as prohibited rather than scoped as high-risk. For a genuine point 6 recidivism tool it surfaces the public-body obligations: the Article 27 FRIA, Article 14 human oversight, and the Article 49 registration into the restricted database section. The engine is deterministic and rule-based, with each finding cited to the Article or Annex point that fired.
Related guides
- Crime-prediction AI: prohibited or high-risk?
- Law-enforcement risk-assessment AI under Annex III point 6
- Predictive-policing AI and the EU AI Act
- Prohibited vs high-risk AI: where the line sits
- The Fundamental Rights Impact Assessment (Article 27)
- Prohibited AI practices under Article 5
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 →