Les outils actuariels de prédiction du risque de récidive se sont largement diffusés dans les systèmes pénaux internationaux, où ils sont mobilisés pour éclairer les décisions judiciaires et pénitentiaires. En France, leur usage demeure limité et fortement discuté. Par ailleurs, si l’intelligence artificielle (ci-après IA) n’est pas encore associée à ces outils, elle constitue un horizon technologique proche. Cet article propose une analyse des apports et limites des outils actuariels existants, des résistances à leur utilisation dans le contexte français et des conditions éthiques, méthodologiques et professionnelles nécessaires à une éventuelle intégration future de l’IA dans l’évaluation du risque de récidive.

This article examines the current and prospective role of actuarial risk assessment tools in predicting recidivism, examining their strengths, limitations, and the potential contribution of artificial intelligence (AI) to their scoring processes. While actuarial instruments have been widely implemented in international criminal justice systems, their adoption in France remains limited and contested. In parallel, the rapid development of AI raises new questions regarding its possible integration into risk assessment practices. The objective of this paper is therefore to analyse the empirical and methodological foundations of actuarial tools, identify the professional and structural barriers to their use, and explore both the opportunities and risks associated with the introduction of AI in their application. Actuarial tools offer several well-documented advantages. By relying on statistically derived risk factors, they provide more accurate predictions of recidivism than unstructured clinical judgment, demonstrating moderate to high predictive validity. They contribute to reducing both false positives and false negatives, thereby improving decision-making consistency. Their standardized structure enhances inter-rater reliability and reproducibility and facilitates a more efficient allocation of resources in line with the Risk-Need-Responsivity model. In addition, these tools promote transparency and evidence-based practices, supporting a more equitable approach to risk assessment. However, these benefits are accompanied by several barriers to their use, though partly mitigated in practice. First, actuarial tools estimate probabilities based on group data and cannot predict individual behaviour with certainty, raising concerns about their use in individualized judicial decisions. However, they are not designed to replace decision-making but to inform it, providing structured probabilistic estimates that must be interpreted alongside case-specific information. Their applicability may also be affected by the origin of the samples on which they were developed, often North American populations, which can limit their transferability to other legal and cultural contexts. Yet, relative risk estimates have shown greater stability across samples, and ongoing efforts aim to recalibrate and adapt these tools to local populations. Furthermore, these tools rely primarily on static factors and do not adequately capture dynamic changes in individuals’ trajectories, which may reduce their relevance for ongoing supervision and intervention planning. This issue is commonly addressed by combining them with dynamic risk assessment instruments and clinical evaluation. Another challenge lies in the potential performative effect of risk scores, which may influence professional practices and lead to an overreliance on quantitative outputs at the expense of qualitative analysis. However, guidelines emphasize that actuarial results should remain one component within a broader, multidimensional assessment. Finally, resistance among practitioners, particularly in the French context, reflects broader concerns about the “technicization” of professional judgment, the preservation of individualized approaches, and the risk of transforming decision-support tools into prescriptive instruments. Nonetheless, training, progressive implementation, and framing these tools as decision-support aids rather than substitutes for professional judgment can facilitate their acceptance. The integration of AI into actuarial assessment processes presents both promising opportunities and significant challenges. On the one hand, AI could enhance the standardization and efficiency of scoring by automating the extraction and coding of relevant information from large and complex judicial or penitentiary records. Natural language processing techniques may reduce inter-rater variability, improve reproducibility, and decrease the time required for assessments. In this sense, AI could serve as a valuable decision-support tool, allowing professionals to focus more on qualitative interpretation and contextualization. On the other hand, the use of AI introduces critical limitations and ethical concerns. Automated systems may generate coding errors due to misinterpretation of ambiguous or incomplete data and lack the contextual understanding required for complex cases. More fundamentally, AI models trained on historical data may reproduce or amplify existing biases within the criminal justice system, including disparities in prosecution, sentencing, and incarceration across social groups. The delegation of certain tasks to automated systems also raises the risk of overreliance on algorithmic outputs, potentially leading to cognitive bias and a reduction of complex situations to simplified risk scores. In addition, the processing of sensitive judicial and medical data requires strict safeguards to ensure data protection, transparency, and accountability, in line with emerging regulatory frameworks such as the European AI Act. In conclusion, actuarial tools and AI should neither be viewed as neutral solutions nor as inherent threats. Their relevance depends on the conditions under which they are used. AI may enhance the reliability and consistency of actuarial scoring, but it must remain strictly supervised and integrated within a broader evaluative framework that includes dynamic factors, contextual analysis, and professional judgment. A cautious, transparent, and ethically grounded approach is therefore essential to ensure that these technologies support, rather than undermine, fair and individualized decision-making in the criminal justice system.