tel-05498240 Predictive Artificial Intelligence Tools in the U.S. Criminal Trial : Reflections (…)

Under the aegis of efficiency and neutrality, algorithmic risk-assessment tools have been rapidly inserted into the realm of pretrial justice, promising scientific and objective prediction where presumption of innocence and scarce resources are pitted against each other. Their sprawl has been uneven and confined to looser pretrial rules than would govern similar proof at trial. These systems translate miscellaneous data into scores of “risk,” inviting courts to act on what might happen rather than on what has been proven, within a process already plagued by plea bargaining, racial disparity, and institutional asymmetry. This thesis argues that such tools do not merely constitute information about the defendant; they reorder what counts as usable knowledge and who is treated as a credible knower at the weakest stage of a criminal process. The thesis lays out how scores interact with judicial decision-making; how ostensibly neutral inputs (criminal history, prior contacts, residence, employment) can re-encode social disadvantage; and how opacity and trade-secret claims frustrate fundamental rights such as due process. The doctrinal analysis clarifies the boundary between regulation and punishment under United States v. Salerno, the demands of proportionality, and the significance of the presumption of innocence when decisions are predictive. A brief technical account explains why claims of “explainability” often exceed what such models can reliably demonstrate for legal purposes, but the core argument is juridical: claims of legitimacy must be substantiated by enforceable safeguards; falling short of that, systems described as ‘explainable' or ‘fair' remain as fancy adjectives. The thesis contends that the safeguards that govern the trial must be applied earlier in the criminal process, where liberty is already fragile. Accordingly, the thesis argues for invoking evidentiary rules early on: demonstrating reliability of the tools; reasoned judicial decisions that explain how any score was weighed against case-specific facts; meaningful opportunities for defence challenge; and, where appropriate, access to independent expertise to translate model outputs without displacing judicial authorship. The researcher is of the opinion that these emphases are necessary so that pretrial decisions remain individualised and proportionate and detention retains its status as an exceptional, non-punitive measure. Finally, it defends doubt as a juridical virtue, a disciplined moment of pause that reins in the allure of algorithmic certainty and reasserts judicial responsibility, so that the presumption of innocence and liberty remains the rule even in a predictive age.