Oxford-McGill study: AI bias in COMPAS is amplified, fairness is free

A new paper from University of Oxford's Ignacio Cofone and McGill University's Warut Khern-am-nuai challenges a central assumption in algorithmic fairness: that making AI fairer inevitably reduces its predictive accuracy. Using causal inference methods on the COMPAS dataset—a widely studied tool in criminal justice risk assessment—the authors show that racial bias is not only present but is exacerbated by models trained on it. Widely used models do more than replicate human bias; they actively deepen it, undercutting claims that algorithmic decision-making is a neutral improvement over human judgment.

The paper also reframes the fairness-accuracy tradeoff. The authors argue that unconstrained models are often treated as an optimal baseline, but those models operationalize 'risk' through flawed normative choices about what to predict and how. In criminal justice, rearrest data—used as the outcome variable—captures and magnifies systemic racial disparities. Applying fairness constraints can correct these distortions without sacrificing accuracy. Published in the Indiana Law Journal (48 pages, 12 tables), the findings extend beyond criminal justice to lending, hiring, and housing, where biased outcome variables reinforce inequality independently of proxy selection. The authors call for law and policy to treat fairness adjustments as cost-effective corrections, not costly tradeoffs.