What Is Intersectional Bias?
Intersectional bias occurs when discrimination targets individuals at the intersection of multiple protected characteristics. An AI tool might treat women fairly overall and treat Black candidates fairly overall, but significantly disadvantage Black women specifically.
Why Standard Audits Miss It
Traditional bias audits analyze one protected characteristic at a time: race, then gender, then age. This approach assumes bias operates independently along each axis. But discrimination often compounds: the experience of a Latina woman in hiring is not simply the sum of biases against Latinos plus biases against women.
Real-World Examples
Research has documented cases where:
- AI resume screeners showed no gender bias and no racial bias individually, but significantly disadvantaged women of color
- Automated interview tools rated White men and Black women similarly to the average, but showed large disparities for Black men and Asian women
- Scoring algorithms produced fair outcomes by race and by age separately, but severely disadvantaged older women
How to Detect Intersectional Bias
- Cross-tabulate outcomes: Examine selection rates for combinations of protected characteristics (e.g., Black women, Asian men, White non-binary)
- Apply statistical tests: Use Fisher's exact test for small intersectional subgroups
- Compare compound groups: Check whether any intersectional group falls below the four-fifths threshold relative to the highest-performing intersectional group
- Visualize distributions: Plot score distributions for intersectional subgroups to identify patterns
The Legal Dimension
Courts have recognized intersectional discrimination claims since the landmark DeGraffenreid v. General Motors case. The EEOC considers intersectional bias in enforcement actions. Employers who only audit single-axis bias may still face liability for intersectional discrimination.
How OnHirely Handles Intersectional Bias
OnHirely's Pro plan includes full intersectional bias analysis. We automatically cross-tabulate outcomes across race, gender, and age categories, apply appropriate statistical tests for each subgroup size, and flag intersectional disparities that single-axis analysis would miss.