Intersectionality and Compounding Bias
What you'll learn: How bias accumulates across multiple identity dimensions, creating disproportionate harm for people with intersecting marginalized identities.
The Core Concept
Intersectionality recognizes that people hold multiple identities simultaneously—race, gender, age, disability status, socioeconomic background, and more. In AI systems, bias doesn't affect each identity dimension independently. Instead, biases compound and interact, creating unique disadvantages that are worse than the sum of individual biases.
For example, an AI hiring system might perform adequately for white women and Black men separately, but fail dramatically for Black women—a group experiencing both gender and racial bias simultaneously. The system might associate "professional" more strongly with white features and male names, doubly penalizing candidates at the intersection.
Why This Matters
When you measure bias along single dimensions (just gender, just race), you miss these compounding effects. A model might pass fairness tests for each protected attribute individually while still causing severe harm to intersectional groups.
The Multiplication Effect
Think of bias like discount penalties at checkout: a 10% gender penalty and 10% race penalty don't simply add to 20%—they multiply. You lose 10% of 90%, resulting in 19% total loss, and the effect worsens with each additional dimension.
Practical Implications
- Disaggregate your metrics: Don't just measure "gender bias" and "race bias" separately
- Create intersectional test cases: Explicitly test combinations like "elderly disabled women" or "young transgender people of color"
- Small sample challenge: Intersectional groups are often underrepresented in datasets, making both training and evaluation harder
Key Takeaway: Bias compounds across identity dimensions in non-linear ways—evaluating fairness requires explicitly testing intersectional groups, not just single attributes in isolation.