Course contentsShow
← AI Engineering
Lesson 1573 of 1,88638. Bias, Fairness, and AlignmentFree lesson

Intersectionality and Multi-attribute Fairness

Evaluating fairness across combinations of protected attributes like race and gender.

Intersectionality and Multi-attribute Fairness

What you'll learn: How to evaluate AI fairness when multiple protected attributes combine in ways that create unique patterns of bias.

The Core Challenge

You've learned to measure fairness across single attributes like race or gender separately. But real people exist at the intersection of multiple identities—a Black woman experiences the world differently than either white women or Black men. Evaluating fairness on race alone, then gender alone, can miss bias that emerges specifically at these intersections.

Intersectional fairness means examining AI system performance across combinations of protected attributes simultaneously, not just in isolation.

Why Single-Attribute Analysis Falls Short

Imagine an LLM resume screener that shows:

  • Equal accuracy for men vs women (passes gender fairness)
  • Equal accuracy for white vs Black candidates (passes race fairness)

Sounds fair, right? But when you check the intersection, you might discover it severely underrates Black women specifically—a harm invisible when analyzing each attribute separately. The model could favor white women and Black men while discriminating against the intersection.

The Combinatorial Challenge

With even three binary attributes (race, gender, age), you have 8 subgroups to analyze. Real systems involve dozens of protected characteristics with multiple categories each—the number of intersections explodes. You must prioritize which intersections matter most for your application context and stakeholders.

Multi-attribute fairness requires disaggregated metrics: compute your fairness measures (demographic parity, equalized odds, etc.) not just for broad groups, but for granular intersectional subgroups.

Key Takeaway: True fairness requires examining how protected attributes combine—bias often hides at intersections that single-attribute analysis cannot detect, demanding explicit evaluation of multi-attribute subgroups.