When Interpretability Is Unequally Distributed: Fairness in Hybrid Interpretable Models

Hybrid interpretable models create a new fairness problem: These systems combine transparent decision-making with black-box AI by routing some cases to each component. The paper identifies that certain demographic groups may systematically receive explainable decisions while others are disproportionately sent to the unexplainable black box—a procedural fairness issue distinct from traditional accuracy-based fairness concerns.

Interpretability disparities are substantial in real-world settings: Testing across multiple datasets and demographic attributes, the researchers found significant "Interpretability Coverage Disparity" (ICD)—meaning different groups don't receive equal access to interpretable decisions—particularly when both the transparent and black-box components are actively being used.

The problem is fixable without major accuracy tradeoffs: Simple constraints that enforce equal interpretability coverage across demographic groups substantially reduced disparities while maintaining model accuracy and sparsity (simplicity), suggesting that fairness and performance need not be in fundamental conflict.

Current auditing practices are incomplete: The paper argues that hybrid interpretable models should be evaluated not just for whether their predictions are fair, but also for whether they fairly distribute interpretability itself—a dimension that standard fairness audits typically overlook.

Fairness improvements can be a co-benefit: In several cases, addressing interpretability disparities also improved standard algorithmic fairness metrics, suggesting that ensuring equal access to explanations may help address other fairness concerns simultaneously.