Risk is non-binary: Re-Evaluating Offender Risk Assessment Instruments

Offender Risk Assessment Instruments (ORAIs) have shaped criminal justice decisions for over three decades, yet their predictive reliability remains contested. This article critiques the statistical architecture underlying widely used tools—particularly their reliance on additive models that neglect interaction effects between variables such as age, prior convictions, and socio-economic context. Drawing on recent literature and case studies (including COMPAS, OGRS, and ARMS), it argues that most ORAIs reduce complex human behaviours to binary outcomes, misclassifying risk and reinforcing systemic bias. The paper introduces the concept of interaction effects, supported by statistical and psychological research, to illustrate how risk is often conditional rather than cumulative. It further explores the limitations of regression models and proposes a range of alternatives, including multi-level modelling, Bayesian updating, Explainable AI, and co-produced desistance-informed profiles. Beyond improving tools, the article calls for a paradigm shift—from control to collaboration, from prediction to understanding. Risk, it concludes, should be treated not as a fixed score but as a dynamic, contextual process. Only by recognising this can justice systems promote accurate assessments, reduce harm, and rebuild public trust.