Can We Predict Crime? Should We?

Prison sentences serve various purposes, including the incapacitation of imprisoned individuals, i.e., their removal from society to prevent further crimes, as well as their rehabilitation, i.e., modifying their behavior so that the probability of committing a new crime is reduced.Early release from prison under parole creates a tension between these two goals, as modern justice systems recognize that it must be subject to scrutiny due to the risk of new crimes, but that in general it reduces the probability of recidivism because during probation individuals can receive support (e.g., for finding a job), and have a strong incentive for avoiding criminogenic behaviors such as abusing alcohol or other substances.Parole decisions taken by judges or parole boards depend on the assessment of the risk of recidivism. The main question is whether the probability of committing a new crime is low enough to support the rehabilitation goal. For the affected individual and for society, there are fewer decisions with more at stake than this one.Many researchers on algorithmic fairness have been brought to the field motivated by the analysis of decision support systems to predict recidivism risk. In Catalonia, a recidivism risk assessment tool developed using historical data about past releases from prisons has been in use for about 15 years. It does not take decisions autonomously, but it is the centerpiece of the process that leads to a petition of parole. According to our research, it is comparable in accuracy to human experts, useful for making consistent decisions, well regarded by prison officials, and it tends to be more lenient than them. However, it is also categorized as a high-risk application by the new AI Act and subject to stringent safeguards that the current tool does not comply with. Evaluating how human intelligence can be complemented with data-driven decision support systems in an ethical and lawful way in this scenario remains a challenging task, and serves as a model for other risk assessment processes done by public administrations in high-stakes situations including child neglect and homelessness.