Justification, Decision Thresholds, and Randomness

Abstract: Procedurally fair decision-making in the fair machine learning literature is primarily understood in terms of a requirement of equal treatment, or treating like cases alike.

In the fair machine learning literature, equal treatment is understood as requiring at least the following two conditions: (1) the same threshold is applied regardless of one’s social identity or arbitrary characteristics, and (2) some form of parity in error rates, regardless of one’s social identity or arbitrary characteristics.

Criticisms of the fair machine learning literature mainly focus on (2) (e.g., Eva 2022). In this talk, I focus on (1), or the application of the same threshold. I argue that thresholds violate a plausible further notion of fairness, that of respecting claims in proportion to their strength (Broome 1991). This account of fairness pushes us towards the greater use of (weighted) lotteries for algorithmic decision-making.