Statistics Seminar

Abstract: Much of the literature on fair machine learning and algorithmic justice has focused on optimizing decision quality. However, this paper argues that assessing the quality of algorithmic decision outcomes is important but insufficient for assessing algorithmic justice in a sufficiently comprehensive way. In addition, we must critically scrutinize our decision landscape: which alternative decision outcomes are available? Which alternative decision problems could, and should, we be solving with the help of algorithmic models?

Using conceptual and normative resources from analytic political and moral philosophy, this paper shows that considering decision landscapes in conjunction with decision quality yields important and useful insights into the question of whether we ought to 

deploy

 a given algorithmic tool in a given domain at all. More specifically, the decision landscapes framework suggests that if optimizing for decision quality in problem P1 moves us further away from key normative goals like justice, we should solve a different problem P2 instead.