Fairness with respect to Stereotype Predictors: Impossibilities and Best Practices

As AI systems increasingly influence decision-making from consumer recommendations to educational opportunities, their accountability becomes paramount. This need for oversight has driven extensive research into algorithmic fairness, a body of work that has examined both allocative and representational harms. However, numerous works examining representational harms such as stereotypes encompass many different concepts measured by different criteria, yielding many, potentially conflicting, characterizations of harm. The abundance of measurement approaches makes the mitigation of stereotypes in downstream machine learning models highly challenging. Our work introduces and unifies a broad class of auditors through the framework of stereotype predictors. We map notions of fairness with respect to these predictors to existing notions of group fairness. We give guidance, with theoretical foundations, for selecting one or a set of stereotype predictors and provide algorithms for achieving fairness with respect to stereotype predictors under various fairness notions. We demonstrate the effectiveness of our algorithms with different stereotype predictors in two empirical case studies.