Algorithmic bias: Looking beyond data bias to ensure algorithmic accountability and equity

Abstract

As algorithms increasingly aid public sector decision making in the United States, it becomes important to understand how to effectively tackle algorithmic bias in systems that local, state, and federal government entities use and procure, including what kinds of policies are currently in place or proposed. There is a prevalent belief that algorithmic bias arises primarily from statistical biases present in the data used to train or develop the algorithm, but bias may also arise during data collection, problem specification, how and where algorithms are deployed, and within the broader societal contextualization of algorithms. So far, enacted policies in in the US center on temporary bans of particular types of algorithms, transparency, and post-hoc bias audits, as well as more wide-ranging (but non-binding) policy frameworks; all largely focus on quantitative notions of fairness when they assess bias, leaving room for more comprehensive legislation to meaningfully address this issue going forward.