A team of researchers has built a way to train AI models that catches bias hiding at the intersections of protected groups, not just within single categories.
The paper, posted to arXiv, proposes a framework that uses mixed-integer optimization to train classifiers that are both interpretable and intersectionally fair. The authors prove that two existing fairness measures, MSD and SPSF, are mathematically equivalent when it comes to identifying the most unfair subgroup. They then show their optimization-based algorithm finds that bias more effectively than prior methods. The result is a classifier that keeps intersectional bias below a set threshold while staying interpretable, rather than treating fairness as a black-box add-on.
Most bias audits still check one attribute at a time, say race or gender, and miss discrimination that only shows up when categories combine, like outcomes for a specific race-and-gender pairing. Regulations like the EU's AI Act require bias mitigation but don't define what counts as bias, leaving a gap this kind of mathematically bounded, interpretable method could fill for finance and healthcare systems under compliance pressure.
It's a narrow academic result, not a product, so whether regulators or banks actually adopt mixed-integer optimization over simpler fairness checks remains the open question.