Pretrial Risk Assessment Algorithms in the Courtroom

Abstract

This dissertation investigates how algorithmic risk assessments used in criminal courts shape decision outcomes, courtroom debate, and the claims that judges, prosecutors, and defense attorneys advance about defendants. Focusing on a tool intended to help judges set pretrial release conditions, I build on a pre-existing randomized controlled trial in a U.S. county that randomized whether the risk assessment report for a given arrested individual was provided to the court or withheld. Using a mix of qualitative and quantitative methods, I analyze administrative data and a sample of court hearings transcripts from the county. Chapter 2 explores the effect of access to risk assessment reports on how often judges order cash bail, with a focus on how the information influences whether prosecutors and defense attorneys secure their requested bail conditions. The results indicate that risk assessments can have an asymmetric effect, with recommendations for cash bail exerting more influence than recommendations to not require cash bail. Chapter 3 identifies two additional ways that the risk assessment tool altered the behavior of judges: how often they required that defendants be subject to pretrial supervision and how they spoke to defendants. Chapter 4 tests theoretical predictions that risk assessments might affect how defendants are evaluated and shape courtroom actors' understandings of the goals of their decision-making. It demonstrates that risk assessment tools may not lead courtroom actors to fundamentally reorient their approach to evaluating defendants, but may spur shifts in some specific domains of their practices. Some of these shifts may be regarded as counterproductive, while others may help align practices with legal ideals. Together, the dissertation contributes to our understanding of the on-the-ground impact of algorithmic decision aids and highlights some of the trade-offs that the adoption of such tools may entail.