Intersectional Fairness

Intersectional fairness means AI systems should treat people fairly even when they belong to multiple marginalized groups at once—like being both a woman and a person of color, or being elderly and disabled. Check this: When evaluating AI systems, ask whether they were tested for fairness across combinations of identities, not just one characteristic at a time.

Intersectional fairness addresses a critical blind spot in AI fairness: treating demographic categories as independent when real people experience overlapping identities simultaneously. Standard fairness metrics (like demographic parity or equalized odds) are typically computed separately for race, gender, age, or disability status. This assumes fairness for each group individually guarantees fairness for combinations—but research shows this is mathematically false. Subgroups at intersections can experience amplified harms that remain invisible in aggregate metrics.

Consider a facial recognition system tested for fairness across racial groups (white, Black, Asian) and across gender (male, female) separately. Both analyses might show acceptable error rates. Yet Black women—who sit at the intersection—often face dramatically higher error rates because the system wasn't trained on enough examples of their specific intersection. The harm compounds rather than averaging out. This pattern repeats across domains: hiring algorithms that discriminate against women with disabilities, credit scoring systems that penalize immigrants of color, healthcare AI that underestimates pain for elderly Black patients.

The practical challenge is combinatorial explosion. Most organizations lack representative data for all these combinations. This forces tradeoffs: prioritize certain intersections based on risk assessments, use synthetic data or reweighting techniques, or accept that some harms will go undetected. The emerging best practice is "intersectional auditing"—proactively testing highest-risk intersections and designing systems that fail gracefully when data is sparse rather than silently degrading.

Intersectionality, coined by legal scholar Kimberlé Crenshaw (1989), describes how overlapping social identities create combined systems of discrimination. In AI ethics, this framework reveals why single-axis fairness metrics fail to protect multiply-marginalized groups. Research demonstrates that subgroup fairness guarantees are mathematically independent from group-level fairness—satisfying the latter does not ensure the former.

• Mitchell et al. (2019): Algorithmic fairness evaluated only on protected attributes independently can mask severe disparities at intersections of multiple attributes; they introduce "intersectional fairness" as requiring that no intersectional subgroup experience worse outcomes than the population average. [1]
• Kearns et al. (2018): Develop "multicalibration" techniques to ensure predictions are well-calibrated not just overall but across exponentially many intersecting subgroups, providing a computationally tractable approach to intersectional fairness. [2]
• Buolamwini and Gebru (2018): Found that commercial gender classification systems performed nearly perfectly on lighter-skinned males (error rate 0.8%) but failed dramatically on darker-skinned females (error rate 34.7%), demonstrating how intersectional identity compounds algorithmic harm. [3]

Key formal concept: subgroup fairness requires that a model's error rate or prediction distribution satisfies fairness constraints not only for major demographic groups but for all combinations of protected attributes. This is stricter than group fairness and requires specialized auditing techniques.

Intersectional fairness faces fundamental practical and theoretical challenges. Data sparsity is the primary barrier: many intersectional groups have small sample sizes, making statistical validation difficult and potentially violating privacy constraints. The combinatorial explosion of possible identity combinations means comprehensive testing is impossible—researchers must prioritize which intersections to audit, which itself involves value judgments about which harms matter most. Some scholars argue that intersectional fairness, while necessary, remains insufficient without addressing root causes of bias in training data and institutional structures. Others note that focusing on technical fixes within AI systems may distract from broader societal inequalities these systems often mirror and amplify.