Rethinking Technical Algorithmic Fairness

Abstract: Artificial intelligence can exhibit favoritism toward certain groups or individuals over others, which may amplify social inequities. Various technical bias identification and mitigation strategies have been proposed for different model development stages to address the problem. A common fairness criterion is statistical independence between sensitive variables (or their proxies), such as demographics and outcome variables. However, there is an ongoing debate about the efficacy of these techniques and the conditions under which they are most effective. In addition, the literature also exhibits conflicting technical formalizations of fairness and the prioritization of performance and generalizability values in the algorithmic fairness community, despite its primary mission of promoting moral values. Should the ethics of high-stakes AI systems, in particular, fairness, be left only in the hands of the developers? Drawing parallels with the Kidney Allocation System, we argue that group and individual voices should be accounted for to fully encapsulate the concept of fairness and effectively mitigate bias. To this end, we propose a machine learning framework wherein multiple stakeholders, including computer scientists, social scientists, communities, and policymakers, collaboratively define fairness criteria tailored to the specific needs of groups or individuals. Upon deployment of AI models, these defined criteria are translated into technical constraints to refine the models. Unlike traditional technical solutions, the proposed framework mandates a collaborative, multistakeholder approach to solving intricate moral dilemmas such as fairness.

Russ Altman – Stanford University; Lea Goetz – GlaxoSmithKline