ARTIFICIAL INTELLIGENCE
🌎 Addressing Bias from the Intersection of Identity
src:gpt
An AI system can be perfectly fair to women. Perfectly fair to Black people. And still systematically discriminate against Black women.
While researching bias this week looking for new updates and discoveries on algorithmic bias and fairness, I encountered this new research which gave me another perspective on discrimination and biases. Most tools I have reviewed fix bias one group at a time, but this research claimed that fixing bias one dimension at a time is not fixing bias at all. This made me decide to dig deeper.
The research launched in January 2026 by remarkable researchers Jiří Němeček, Mark Kozdoba, Illia Kryvoviaz, Tomáš Pevný, and Jakub Mareček from Czech Technical University in Prague and Technion challenges the foundational assumption of how fairness has been measured and enforced in AI systems.
Just imagine this:
A bank deploys an AI credit scoring system. The compliance team runs a fairness audit. The system passes, no gender bias detected, no racial bias detected. The audit report is filed. The regulators are satisfied. But nobody checked what happened to older Black women applying for small business loans. Their approval rate is 34% lower than any other group. The AI never saw them as a category. The audit never looked. The bias was hiding in the intersection exactly where the tools were never designed to look.
A system declared fair by every available measure was quietly denying a community access to capital.
This research explains exactly why that bank's audit missed the problem — and introduces the framework that would have caught it.
Key Findings
⚖️ Marginal Fairness Is Not Enough: Achieving fairness for single protected groups like gender, race, age can still mask severe discrimination against people who belong to multiple protected groups simultaneously.
🔗 Two Measures Are Actually One: The paper proves mathematically that MSD and SPSF two previously competing measures of intersectional bias identify the same most-unfair subgroup, unifying the field around a single detection approach.
🧠 The MIO Framework: Mixed-Integer Optimization trains AI classifiers to be fair across all intersecting protected groups at once bounding intersectional bias below an acceptable threshold while maintaining high accuracy.
🔍 Interpretability Built In: Unlike black-box fairness tools, MIO identifies the exact subgroup being discriminated against in plain language for example, "female AND white AND age over 40" making bias actionable, not just measurable.
📊 Tested on Real High-Stakes Data: Validated across five US Census datasets covering income, employment, healthcare, mobility, and travel time, the domains where intersectional discrimination causes the most real-world harm.
Why It Matters
For FinTech and Lenders: Credit scoring and loan approval systems that pass single-attribute audits may still discriminate at intersections. This framework provides the first practical tool to catch and correct that hidden bias before regulatory exposure.
For HR Technology and JobTech: Hiring algorithms audited for gender or racial fairness alone remain vulnerable to intersectional discrimination claims. Intersectional auditing is no longer optional it is the next legal frontier.
For Policymakers and Regulators: The EU AI Act mandates bias mitigation but stays deliberately vague on definition. This paper gives regulators a mathematically grounded, interpretable standard to actually enforce.
For Everyone: The most disadvantaged people in any system are rarely disadvantaged in just one way. Any fairness framework that ignores intersectionality is not measuring fairness, it is measuring the absence of the most visible unfairness.
Reviewed & Written by Oluwasegun Odesola | DataIntell
Paper: Read More