Abstract

We argue in this article that the integration of fairness into machine learning, or FairML, is a valuable exemplar of the politics of statistics and their ongoing transformations. Classically, statisticians sought to eliminate any trace of politics from their measurement tools. But data scientists who are developing predictive machines for social applications - are inevitably confronted with the problem of fairness. They thus face two difficult and often distinct types of demands: first, for reliable computational techniques, and second, for transparency, given the constructed, politically situated nature of quantification operations. We begin by socially localizing the formation of FairML as a field of research and describing the associated epistemological framework. We then examine how researchers simultaneously think the mathematical and social construction of approaches to machine learning, following controversies around fairness metrics and their status. Thirdly and finally, we show that FairML approaches tend towards a specific form of objectivity, "trained judgement," which is based on a reasonably partial justification from the designer of the machine - which itself comes to be politically situated as a result.