Algorithmic processing and AI bias: using overfitting to reveal rather than perpetuate existing bias.

In this paper we analyse AI overfitting in algorithmic processing to show how it relates to cases of unfairness or AI bias and how it combines with complex social phenomena such as looping effects to maintain and exacerbate existing bias. We discuss existing and proposed AI regulation attempting to address this bias to pick up dominant trends and priorities. Finally, we suggest that, although the focus of the literature currently falls on the negative consequences of overfitting, it can be used as a diagnostic tool for detecting underlying social inequalities and, as such, lead to alternative uses of AI analytics to expose social injustice rather than exacerbate it. This paper provides further theoretical support to recent views in the literature suggesting that algorithmic processing can be used to diagnose and monitor bias; by highlighting the interaction with looping effects, it also provides additional motivation to use overfitting as a first step towards mitigation of historical prejudice.