When are Algorithms Biased? A Multi-Disciplinary Survey

Abstract

Bias means different things to lawyers, psychologists, statisticians, and computer scientists. While such interdisciplinary variation obfuscates the definition and thus management of algorithmic bias, it may also shed light on the nature of bias itself. This paper surveys bias definitions across these disciplines, analyses their intersection, and assesses how they shape when algorithms are said to be “biased”. It reveals bias as a systematic deviation from a presupposed ideal. To call an algorithm “biased” is to claim that ideal worth pursuing. Where mathematical ideals, like statistical expectations, are put forth, bias assertions plausibly avoid norm assertions. Existing accounts of algorithmic bias, however, go further than the objective sciences and, by asserting that some algorithm or dataset is unfair, implicates the normative arts. This underscores the necessity of multi-disciplinary perspectives and communication in measuring and managing algorithmic bias.