Algorithmic Fairness and Social Welfare

What constitutes a fair algorithm? In the literature on algorithmic fairness, a common approach is to formulate fairness concerns as statistical constraints and to select the most accurate algorithm satisfying this constraint. This approach is facially distinct from a long tradition in economics based on social welfare, where the utilities of different social identities are aggregated from behind a veil of ignorance. We show that the constrained optimization and social welfare approaches can be fundamentally opposed and propose a framework that nests both approaches as special cases.