“The Overstated Cost of AI Fairness in Criminal Justice”

The title of this post is the title of this new paper now on SSRN authored by Ignacio Cofone and Warut Khern-am-nuai. Here is its abstract:

The dominant critique of fairness constraints in AI decision-making, particularly in criminal justice, is that they come at a cost: increasing fairness reduces the accuracy of predictions, thereby imposing a social burden. This article challenges that assumption by empirically analyzing the Correctional Offender Management Profiling for Alternative Sanctions (COMPAS) algorithm, a widely used risk assessment tool in the U.S. criminal justice system.

This study makes two key contributions. First, it demonstrates that widely used AI models do more than replicate existing biases—they exacerbate them. Using causal inference methods, we show that racial bias in COMPAS is not only present in its training data but is amplified by common AI algorithms themselves. This finding has implications for legal scholarship and policymaking, as it challenges the assumption that AI can offer an objective or neutral improvement over human decision-making. It also provides counter-evidence to the idea that AI merely mirrors pre-existing human biases.

Second, it reframes the debate over the cost of fairness in algorithmic decision-making. The paper how AI systems used in criminal justice operationalize concepts such as risk and fairness by making implicit and often flawed normative choices about what to predict and how to predict it. The claim that fair AI models decrease accuracy assumes that the model’s predictions are an optimal baseline. Our findings reveal that applying fairness constraints does not necessarily lead to a meaningful loss in predictive accuracy regarding recidivism. Instead, it corrects for the algorithm’s reliance on biased proxies, such as rearrest data, which reflect systemic racial disparities in law enforcement rather than actual criminal risk. In other words, fairness constraints do not degrade the accuracy of algorithms trained to predict recidivism on arrest data; rather, they correct distortions introduced by biased training data and flawed output variables. This distinction is critical for legal debates on AI governance, as it reveals that, in some cases, regulatory interventions can introduce algorithmic fairness without imposing the tradeoffs often presumed in policy discussions.

These implications of these findings extend beyond criminal justice. Similar dynamics exist in AI-driven decision-making in lending, hiring, and housing, where biased output variables reinforce systemic inequalities beyond the choices of proxies for desirable traits. By providing empirical evidence that fairness constraints can improve rather than undermine decision-making, this article advances the conversation on how law and policy should approach AI bias, particularly in contexts where algorithmic decisions affect fundamental rights.