Intersectional Fairness

The foundational process of defining and isolating intersectional subgroups by cross-referencing multiple protected attributes. A standard group-fairness audit might check for bias against 'women' and 'Black individuals' separately, but fail to detect severe discrimination against the specific subgroup of 'Black women.' Intersectional fairness requires explicitly creating these compound categories—such as 'disabled, elderly, non-native speakers'—to ensure no combination of attributes experiences a uniquely adverse outcome that is statistically invisible in aggregate metrics.

A core problem motivating intersectional fairness, where a model appears fair on every individual protected attribute but is deeply unfair to specific intersectional subgroups. An algorithm can satisfy both gender parity and racial parity simultaneously while systematically denying opportunities to women of a specific race. This occurs because aggregate group metrics can be mathematically manipulated, or 'gerrymandered,' to hide harm concentrated at the intersections. Detecting this requires auditing a combinatorially large set of subgroups, making computational efficiency a central challenge.

A powerful technical framework for achieving intersectional fairness. A predictor is multicalibrated if it is simultaneously calibrated not just on the overall population, but on every subgroup identifiable by a chosen class of computationally bounded auditors. This provides a strong, provable guarantee: for every intersectional subgroup an auditor can define, the model's predicted probabilities accurately reflect the true empirical likelihood of the outcome. Multicalibration inherently prevents fairness gerrymandering and ensures no identifiable subgroup is systematically mis-calibrated.

The primary computational challenge in intersectional fairness. The number of potential subgroups grows exponentially with the number of protected attributes and their values. Auditing a model with just 10 binary attributes yields over 1,000 distinct subgroups, many with vanishingly small sample sizes. This sparsity makes statistically significant bias detection difficult. Solutions involve:

• Rich subgroup classes: Auditing computationally identifiable groups rather than pre-enumerating them.
• Statistical priors: Using hierarchical Bayesian models to share strength across related subgroups.
• Differential fairness: Applying privacy-inspired noise to bound the influence of any single attribute combination.

A privacy-inspired intersectional fairness definition that applies the mathematical framework of differential privacy to algorithmic outputs. A model satisfies ε-differential fairness if, for any output, its probability given one intersectional subgroup is within an e^ε multiplicative factor of its probability given any other subgroup. This elegantly handles the combinatorial explosion by providing a uniform, worst-case bound on outcome divergence across all possible attribute combinations simultaneously, without requiring explicit enumeration of every sparse subgroup.

Intersectional fairness directly operationalizes the legal doctrine of intersectionality, coined by Kimberlé Crenshaw, which argues that anti-discrimination law must recognize compound identities. In the landmark case DeGraffenreid v. General Motors, the court failed to recognize discrimination against Black women because the employer hired white women and Black men. Modern AI auditing frameworks, including the EU AI Act's mandate for high-risk systems, implicitly require intersectional analysis to avoid this exact failure mode, moving beyond single-axis fairness checks.