A study finds US judges use algorithmic risk assessments for criminal defendants selectively, instead of wholesale adopting or rejecting algorithms' advice
Lauren Feiner / The Verge :
Context & Ripple Effects
US criminal courts have long used third-party recidivism tools in decisions around bail and sentencing, while concerns about their opacity have persisted; earlier coverage of proprietary risk-assessment use in courts framed the accountability problem.
The new finding adds behavioral evidence to a debate already shaped by reported racial-bias concerns in crime-prediction software and by New Jersey's experience, where a pretrial tool coincided with lower jail populations but persistent disparities.
First-order effects
- The study challenges a simple “judge versus algorithm” framing: algorithmic scores appear to function as one input that judges weigh selectively, rather than as automatically followed recommendations.
- Court administrators, researchers, and tool providers must evaluate how scores are used in practice, not only whether a jurisdiction has adopted a risk-assessment system.
Second-order effects
- Audits of criminal-justice AI may need to examine judge-level and case-level patterns of reliance; aggregate accuracy or adoption metrics can miss where discretion changes outcomes.
- The result strengthens the case for governance measures that make both the tool's recommendation and the human rationale reviewable, particularly where disparities remain a concern.
Third-order effects
- If selective reliance is widespread, accountability will increasingly center on the combined human-algorithm decision process rather than on algorithm performance alone.
- That could shift public-safety AI governance toward monitoring discretion, documentation, and contestability—but the study alone does not establish which safeguards produce fairer results.
The trend: Criminal-justice AI is moving from a narrow debate over whether to deploy risk scores toward scrutiny of how institutional decision-makers incorporate them.