Please rotate your device to landscape mode to view the charts.

Key Findings

Machine Learning Can Reduce Crime While Maintaining Release Rates

Algorithm-based decisions could reduce crime by 24.7% with no change in jailing rates, or reduce jail populations by 41.9% with no increase in crime rates.

Judges Miss High-Risk Cases

The riskiest 1% of defendants who judges release have a 56.3% failure-to-appear rate but are released 48.5% of the time.

Algorithm Reduces Racial Disparities

The algorithm can achieve the same crime reduction as judges while jailing 40.8% fewer minorities, demonstrating potential for both improved outcomes and increased equity.

Crime Reduction Potential Across Judge Leniency Quintiles

  • Stricter judges (5th quintile) achieve 30.7% crime reduction but require 22.3% more jailing
  • Algorithm could achieve same crime reduction with only 11.2% more jailing
  • Shows efficiency gains possible through algorithmic decision-making

Racial Impact of Detention Decisions

  • Algorithm achieves same crime rate while detaining 38.8% fewer Black defendants
  • Hispanic detention rates reduced by 44.6%
  • Demonstrates algorithm can improve both efficiency and equity

Risk Assessment Accuracy

  • Algorithm identifies highest-risk 1% with 56.3% failure rate
  • Judges release 48.5% of these high-risk defendants
  • Shows potential for improved risk assessment through machine learning

Contribution and Implications

  • Demonstrates practical value of machine learning in improving judicial decisions while reducing disparities
  • Provides framework for evaluating algorithmic decision aids in presence of selective labels and multiple objectives
  • Shows importance of considering both prediction accuracy and decision outcomes when implementing algorithmic tools

Data Sources

  • Crime reduction comparison chart based on Table III showing relative performance across judge leniency quintiles
  • Racial impact visualization derived from Table VII comparing detention rates by race between judges and algorithm
  • Risk assessment accuracy chart constructed using data from Table II on algorithm's predictive performance