A Survey on Bias and Fairness in Machine Learning

This survey addresses fairness in artificial intelligence, motivated by the widespread deployment of AI systems that make important, sometimes life-changing decisions in sensitive environments. The authors investigate a range of real-world applications that have exhibited bias in various ways and enumerate the different sources of bias that can affect AI applications. To help researchers reason about avoiding such bias, they create a taxonomy of the fairness definitions that machine learning researchers have proposed.

The survey then examines many AI domains and subdomains, describing the unfair outcomes researchers have observed in state-of-the-art methods and the ways they have tried to address them. The authors note that many future directions and solutions remain for mitigating bias in AI systems, and they frame the survey as a way to motivate researchers to tackle these issues by drawing on existing work in their respective fields. As AI moved into commercial and high-stakes use, consolidating the definitions and sources of bias mattered for building fairer systems.