Intersectional Fairness Score : the overlooked but far-reaching choice of aggregation design

EurIPS 2025 Workshop UPLB Submission27 Authors

Keywords: fairness in AI, intersectionality, bias assessment, bias mitigation, fairness score, multi-dimensional aggregation

TL;DR: how to aggregate multiple bias measures among demographic subgroups arising from intersectionality in one single fairness score: the various design possibilities and their inconsistent implications.

Track: Regular Paper

Abstract: Fairness assessment in AI is essential for building responsible models. Traditionally, it focuses on two demographic groups situations, but real-world complexity requires considering intersectionality. This work explores how to aggregate bias measurements across multiple subgroups into one single score—a critical step often overlooked. We first show that the choices made in aggregation design (norm, maximum, probabilistic approaches, etc.) can significantly influence results, leading to divergent or even conflicting conclusions. We identify and analyze the various possible methods, highlighting their ethical implications and providing a first framework of criteria to guide their selection based on context. Our goal is to foster interdisciplinary discussion on this often-neglected step, aiming for a fairer, more informed and transparent evaluation of intersectional biases in AI.

Submission Number: 27