Does the End Justify the Means? On the Moral Justification of Fairness-Aware Machine Learning

Abstract

Despite an abundance of fairness-aware machine learning (fair-ml) algorithms, the moral justification of how these algorithms enforce fairness metrics is largely unexplored. The goal of this paper is to elicit the moral implications of a fair-ml algorithm. To this end, we first consider the moral justification of the fairness metrics for which the algorithm optimizes. We present an extension of previous work to arrive at three propositions that can justify the fairness metrics. Different from previous work, our extension highlights that the consequences of predicted outcomes are important for judging fairness. We draw from the extended framework and empirical ethics to identify moral implications of the fair-ml algorithm. We focus on the two optimization strategies inherent to the algorithm: group-specific decision thresholds and randomized decision thresholds. We argue that the justification of the algorithm can differ depending on one's assumptions about the (social) context in which the algorithm is applied -even if the associated fairness metric is the same. Finally, we sketch paths for future work towards a more complete evaluation of fair-ml algorithms, beyond their direct optimization objectives.

Distributive Justice as the Foundational Premise of Fair ML: Unification, Extension, and Interpretation of Group Fairness Metrics

“…The "leveling down objection" The "leveling down objection" is a prevalent anti-egalitarianism argument [37,14] saying that less inequality is not desirable if this requires lowering the better-off group's welfare to match the one of the worse-off group. On this basis, choosing egalitarianism as the pattern of justice has been criticized in the algorithmic fairness literature (see, e.g., [33,25,49]). Our approach allows using other patterns of justice, such as maximin, prioritarianism, or sufficientarianism (see Section 3.3).…”

“…Existing group fairness criteria thus represent a very confining definition of utility. Our approach acknowledges that the utility of the decision subjects does not only depend on the decision itself but also on other attributes such as one's ability to repay a loan or one's socioeconomic status (see, e.g., [22,49,8]. This is represented through the utility function described in Section 3.1.…”

“…This is what is known as the "leveling down objection", which is often brought forward to challenge egalitarianism in philosophical literature [37,14]: In a case in which equality requires us to worsen the outcomes for everyone, should we really demand equality or should we rather tolerate some inequalities? As criticized by Cooper and Abrams [12], Weerts et al [49], existing definitions of group fairness lack this differentiation as they always minimize inequality.…”

“…No consideration of consequences As pointed out by Hertweck et al [22] and Weerts et al [49], a large part of the existing work on fairness criteria seems to focus on an equal distribution of favorable decisions and not on the consequences of these decisions. Binns [8] notes that these criteria "[assume] a uniform valuation of decision outcomes across different populations" [8, p. 6], and notes that this assumption does not always hold.…”

“…As previously discussed, current definitions of group fairness only consider the decisions themselves, but not their consequences -even though the same decision could be beneficial for some and harmful for others [49]. Our approach explicitly considers the consequences of decisions, i.e., the resulting utility (or welfare), which could be positive in the case of a benefit or negative in the case of a harm.…”