Interventions Against Machine-Assisted Statistical Discrimination

Abstract

I study statistical discrimination driven by verifiable beliefs, such as those generated by machine learning, rather than by humans. When beliefs are verifiable, interventions against statistical discrimination can move beyond simple belief-free designs, like affirmative action and blinding, to more sophisticated belief-contingent ones. I analyze a belief-contingent intervention, common identity, and show that it can be more effective at combating statistical discrimination than popular alternatives -- particularly when the training dataset exhibits the kinds of statistical biases that often plague machine-assisted decision problems.