Predicting Criminal Offending From Personality Traits
The fourth prohibited category covers AI systems that predict a person will commit a crime in the future based on personality traits, psychological profiles, or similar subjective assessments of character. This is sometimes called predictive policing or criminal risk profiling. It is distinct from conventional criminal justice use of prior convictions or specific evidence.
The concern is that such systems are based on unproven science and unreliable foundations. Personality psychology cannot reliably predict individual criminal behaviour, especially when applied to people who have not yet committed crimes. The algorithms often rely on factors that correlate with crime rates in historical data, including poverty, neighbourhood of residence, ethnic background, family structure and prior contact with law enforcement. These factors reflect structural inequality and historic discrimination more than actual predictive power regarding future criminal acts. An AI system trained on historical policing data will inevitably inherit biases from past policing decisions, including discriminatory practices that may no longer be used but left their mark on the data.
Using such a system to make decisions about a person’s liberty, freedom of movement, or access to services is prohibited. This includes decisions to arrest without evidence, to refuse bail, to impose curfew conditions as a condition of bail, to place on a watchlist, or to subject to enhanced surveillance and scrutiny. The harm is clear: an innocent person can be punished or severely restricted based on a statistical prediction that they might commit a crime in the future, even though they have done nothing wrong yet.
The prohibition is absolute and unconditional. Police cannot use an AI system that predicts criminal offending from personality traits, even in the context of a criminal investigation. Neither can private security companies, insurance firms setting premiums, employers making hiring decisions, or government benefits administrators allocating resources. There is no exception for accuracy. Even if the model has high predictive accuracy on the training set, it is prohibited. There is no exception for transparency. Even if the model’s decision is explainable and fully disclosed to the person, it remains prohibited. The prohibition is not about process or clarity; it is about substance.
The law enforcement exception to Article 5 (which permits subliminal and manipulative techniques in criminal investigation) does not extend to this prohibition. Law enforcement cannot use personality-based predictive policing tools. This is one of the few cases where the EU’s prohibition is even stricter for law enforcement than for other sectors. The EU is particularly concerned about state power in this area. When police have the power to predict who will offend based on their characteristics and personality, they have concentration of power that undermines the rule of law.
It is critical to understand that this prohibition does not prevent all AI use in criminal justice. AI can legitimately support police work by analysing specific evidence, identifying patterns in specific cases, validating identifications through biometric systems, or providing information systems to support investigations. What is prohibited is the use of AI to generate risk scores based on who someone is.
This creates a practical challenge for countries that have already deployed such systems. They must decommission them and cease relying on their predictions in any decision-making. It also creates clarity about legitimate AI use in criminal justice. An AI system that uses objective factors directly related to a specific alleged crime is not prohibited. An analysis of specific evidence from a specific investigation is permitted. But a system that assigns a risk score based on personality, psychological characteristics, or inferred propensity to commit crime is banned outright.
The distinction is important and must guide implementation: prediction based on objective factors and specific acts is permitted. Prediction based on who someone is, how they think, or their personality type is prohibited.
Examples of prohibited criminal prediction from personality traits include: a police AI system that assigns a risk score to people based on neighbourhood of residence, family background, prior contact with law enforcement and personality trait inferences, then uses the score to decide who to stop, search or place on a watchlist; a sentencing recommendation system that predicts likelihood of reoffending based on social background, employment history and inferred personality traits, then recommends sentence length; an intelligence AI system that scores communities or neighbourhoods for criminality based on demographic composition and historical crime rates, then directs police resources accordingly. In contrast, an AI system that analyses evidence from a specific investigation, that validates the identity of a suspect using specific biometric or forensic data, or that supports trial decision-making based on facts of the specific case would not be prohibited.