Algorithms As Prosecutors: Lowering Rearrest Rates Without Disparate Impacts and Identifying Defendant Characteristics ‘noisy’ to Human Decision-makers

Daniel Chen (TSE / IAST / Université Toulouse )
Elliott Ash (U. of Warwick)
Daniel Amaranto (NYU)
Lisa Ren (NYU)
Roper Roper (NYU)

Abstract: This paper investigates how machine learning might bring clarity to human decisions made during the criminal justice process. Our data comes from all cases at the New Orleans District Attorney’s office for the years 1988-1999. We exploit random assignment of prosecutors, prosecutorial discretion, and heterogeneity across prosecutors in charge rates to compare prediction models to judicial decision makers. Our model of defendant rearrest, trained using de- fendant and offense characteristics, selects higher-risk individuals to prosecute than its human counterparts did. In particular: given a set charge rate, our model would reduce rearrest rates by five to nine percentage points. This model could have several important policy implications: it might identify defendant characteristics that are particularly ‘noisy’ to prosecutors; it could suggest ways of alleviating criminal caseloads without increasing crime rates; and it might provide important insights into how a prosecutor’s background relates to the quality and nature of their charging decisions.