Lisez cet article à voix haute
Activez le micro et lisez à votre rythme. Les mots s’allument au fil de la lecture, et rien ne vous arrête quand l’un d’eux est mal prononcé.
Votre audio n’est jamais envoyé à MovaReader : c’est votre navigateur qui le reconnaît. Politique de confidentialité
Connectez-vous pour garder cette séance dans votre historique de lecture.
Evaluating Algorithmic Risk Scores in Pretrial Custody Determinations
The integration of statistical risk assessment instruments into pretrial custody hearings reflects an institutional drive toward objective judicial decision-making, yet interpreting algorithmic scores demands disciplined methodological caution. Algorithmic tools generate empirical scores estimating the likelihood that an arrested individual will fail to appear in court or be rearrested before trial. Proponents argue that objective risk classifications can reduce unnecessary pretrial detention while maintaining public safety. However, computational social scientists tend to emphasize that these algorithms do not measure individual criminality directly; rather, they identify correlations within historical criminal justice data. Because arrest records inevitably mirror past policing practices and discretionary enforcement patterns, algorithmic outputs may inadvertently reproduce systemic demographic disparities under the guise of neutral mathematics.
A fundamental theoretical challenge involves mathematically incompatible definitions of algorithmic fairness. Groundbreaking proofs in algorithmic