Fairness audits typically examine sensitive attributes such as race, gender, or age one coordinate at a time. The authors of a new arXiv paper argue that this approach has a blind spot: a score can be independent of each sensitive attribute individually while still depending on their joint distribution. Such intersectional disparities are precisely the kind of harm that fairness tools aim to catch, yet they slip through conventional per-coordinate audits.

The paper introduces a framework that treats multiple sensitive attributes jointly, using Wasserstein distances to quantify distributional fairness. This allows auditors to detect when a model's outputs are unfairly distributed across combinations of attributes, not just along single dimensions. To make the results actionable, the framework also applies Shapley allocation, a concept from cooperative game theory, to decompose any detected disparity into contributions from individual attributes and their intersections.

The work is methodological, offering a way to extend fairness audits beyond univariate checks. By moving from one-dimensional to joint distributional comparisons, it addresses a known limitation in the fairness literature. The authors do not present experimental results in the abstract, so the practical impact on real-world auditing workflows remains to be seen, but the theoretical contribution is a step toward more comprehensive fairness evaluation.