The Overstated Cost of AI Fairness in Criminal Justice

Algorithmic bias in criminal justice is actively amplified, not merely replicated: Using causal inference methods on the COMPAS dataset, the researchers found that models don't just mirror existing human bias—they actually exacerbate racial bias, undermining the assumption that algorithmic decision-making represents a neutral or objective improvement over human judgment.

The fairness-accuracy tradeoff may be a false choice: Contrary to the widespread belief that improving fairness requires sacrificing predictive accuracy, the paper demonstrates that applying fairness constraints to criminal risk assessment models does not necessarily reduce accuracy, challenging a key argument used to justify deploying biased systems.

Biased outcome variables are a hidden source of discrimination: The study reveals that rearrest data used to train these models captures and magnifies systemic racial disparities rather than measuring true criminal risk, meaning the models are optimizing for a fundamentally flawed target that perpetuates inequality independent of how variables are selected.

This problem extends across multiple high-stakes domains: The dynamics identified in criminal justice—where biased historical data distorts model predictions—also apply to lending, hiring, and housing decisions, suggesting the need for broader scrutiny of outcome variables across sectors where algorithmic systems affect people's lives.

Fairness interventions can correct rather than compromise: By reframing fairness constraints as corrections to distorted outcome variables rather than restrictions on accuracy, policymakers can pursue fairer systems without accepting the premise that equity requires sacrificing predictive performance.