Racial Equity in Algorithmic Criminal Justice
68 Pages Posted: 24 Mar 2018 Last revised: 29 Jun 2018
Date Written: June 20, 2018
Algorithmic tools for predicting violence and criminality are being used more and more in policing, bail, and sentencing. Scholarly attention to date has focused on their procedural due process implications. My aim here is to consider these instruments’ interaction with the enduring racial legacies of the criminal justice system There are two competing lenses for evaluating the racial effects of algorithmic criminal justice: constitutional doctrine and emerging technical standards of “algorithmic fairness.” I argue first that constitutional doctrine is poorly suited to the task. It will often fail to capture the full range of racial issues that potentially arise in the use of algorithmic tools in criminal justice. While the emerging technical standards of algorithmic fairness are at least fitted to the specifics of the relevant technology, the technical literature has failed to ask how various conceptions of fairness track (or fail to track) policy-significant consequences. Drawing on the technical literature, I propose a reformulated metric for considering racial equity considerations in algorithmic design. Rather than asking about abstract definitions of fairness, a criminal justice algorithm should be evaluated in terms of its long-term, dynamic effects on racial stratification. The metric of nondiscrimination for the algorithmic context should focus on the net burden placed on a racial minority. Surprisingly, a precise formulation of this metric suggests that it can converge with the socially efficient decision rule under certain conditions.
Keywords: Machine learning; criminal justice; racial equality
Suggested Citation: Suggested Citation