Supposedly Fair Classification Systems and Their Impacts

The algorithmic fairness field has boomed with discrimination mitigation methods to make Machine Learning (ML) model predictions fairer across individuals and groups. However, recent research shows that these measures can sometimes lead to harming the very people Artificial Intelligence practitioners want to uplift. In this paper, we take this research a step further by including real ML models, multiple fairness metrics, and discrimination mitigation methods in our experiments to understand their relationship with the impact on groups being classified. We highlight how carefully selecting a fairness metric is not enough when taking into consideration later effects of a model’s predictions– the ML model, discrimination mitigation method, and domain must be taken into account. Our experiments show that most of the mitigation methods, although they produce “fairer” predictions, actually do not improve the impact for the disadvantaged group, and for those methods that do improve impact, the improvement is minimal. We highlight that using mitigation methods to make models more “fair” can have unintended negative consequences, particularly on groups that are already disadvantaged.