While formal definitions of fairness in machine learning (ML) have been proposed, its place within a broader institutional model of fair decision-making remains ambiguous. In this paper we interpret ML as a tool for revealing when and how measures fail to capture purported constructs of interest, augmenting a given institution's understanding of its own interventions and priorities. Rather than codifying "fair" principles into ML models directly, the use of ML can thus be understood as a form of quality assurance for existing institutions, exposing the epistemic fault lines of their own measurement practices. Drawing from Friedler et al's [2016] recent discussion of representational mappings and previous discussions on the ontology of measurement, we propose a social measurement assurance program (sMAP) in which ML encourages expert deliberation on a given decision-making procedure by examining unanticipated or previously unexamined covariates. As an example, we apply Rawlsian principles of fairness to sMAP and produce a provisional just theory of measurement that would guide the use of ML for achieving fairness in the case of child abuse in Allegheny County.
Read Paper