Algorithmic Sentencing and the Illusion of Neutrality: An Empirical Assessment of Risk-Assessment Tools in Criminal Justice

Actuarial risk-assessment instruments are now embedded in sentencing, bail, and parole decisions across a growing number of jurisdictions. Proponents argue that these tools replace idiosyncratic human judgment with statistically validated, facially neutral criteria. This article interrogates that claim. Drawing on the doctrinal record generated by State v Loomis and subsequent litigation, together with the empirical literature on differential predictive accuracy across demographic groups, we argue that facial neutrality at the level of input variables does not entail neutrality of outcome. We propose a three-part legal framework — disclosure, contestability, and periodic recalibration — through which courts and legislatures can domesticate these tools without abandoning their evidentiary benefits. The article closes by situating the debate within a comparative frame, contrasting the largely common-law, litigation-driven American response with the more anticipatory, regulation-first approach reflected in the European Union’s Artificial Intelligence Act.