Strict Algorithmic Scrutiny: A Formulation of Algorithmic Procedural Fairness Through Affirmative Action Jurisprudence

Abstract

We develop a notion of algorithmic procedural fairness, in both weak and strong forms, inspired by relevant jurisprudence in the realms of discrimination, equal protection, and affirmative action. Formulating affirmative action interventions as a series of potentially stochastic mechanisms governed by a scoring function and a threshold, we develop notions of compound mechanisms which either observe an individual's true group identity with respect to a protected class, or some intermediate representation that are not bound by the same requirements on procedural fairness. Subsequently, we leverage relevant characterizations of Supreme Court treatment of holistic analysis of race to formulate a conception of strongly procedurally fair stochastic mechanisms, noting an interesting compatibility between certain non-negative distributions in the exponential family, including the Exponential Distribution and Geometric Distribution, and probability bounds over arbitrary mechanism threshold. Finally, we connect our work to potential implications in the wider examination of the relationship between algorithmic fairness and the law.