Fairness

Abstract

Abstract Despite widespread agreement that algorithmic bias is a problem, there is a lack of agreement about what to do about it. This chapter argues that what should be done about algorithmic bias depends on whether the problem of algorithmic bias is conceptualized as a problem of fairness, or some other problem of justice. It substantiates this claim by examining the debate over different formal fairness metrics. One compelling metric for measuring whether a system is fair measures whether the system is calibrated, or whether a prediction has equal evidential value regardless of an individual’s group membership. Calibration exemplifies a compelling notion of accuracy, and of fairness, in treating like cases alike. However, there can be a tradeoff between making systems fair, in this sense, and making them more just: to make more accurate predictions, a system may use social patterns that reinforce structures of unjust disadvantage. In response to this tradeoff, the chapter argues that in situations of injustice, other values of justice ought to be privileged over fairness, as fairness has no value in the absence of just background institutions. It concludes by drawing out five proposals for better governance of AI for justice and fairness from the philosophical discussion of fairness and justice in AI. These are a values-first approach to bias interventions, de-coupling decision processes, explicitly modeling structural injustice, interventions to increase data quality, and the use of (weighted) lotteries rather than decision thresholds.