Catalogue of Tools & Metrics for Trustworthy AI

These tools and metrics are designed to help AI actors develop and use trustworthy AI systems and applications that respect human rights and are fair, transparent, explainable, robust, secure and safe.

Interpretability Coverage Disparity (ICD) is a fairness measure for hybrid interpretable machine learning models. It was introduced in the paper "When Interpretability Is Unequally Distributed: Fairness in Hybrid Interpretable Models", published in May 2026 by researchers in Montréal. Hybrid interpretable models combine a transparent component, such as a set of readable rules, with an opaque "black-box" model. Each case is routed to one of the two components. This design lets developers balance accuracy against interpretability. However, it raises a distinct fairness concern: some demographic groups may systematically receive explainable decisions, while others are disproportionately routed to the black box.

ICD measures this imbalance. It compares how often members of different demographic groups are handled by the interpretable component. It depends only on this routing decision, not on the predictions themselves. It is therefore a measure of procedural fairness, capturing unequal access to explainable decisions rather than unequal outcomes. The authors also measure how arbitrary this routing is for individuals across similarly accurate models.

The authors study ICD across four hybrid interpretable learning methods, three standard fairness datasets and several sensitive attributes. They find substantial disparities when both components are actively used. Adding simple constraints on coverage disparity during training significantly reduces ICD, with little loss of accuracy or simplicity, and in several cases also improves standard fairness metrics. The authors conclude that hybrid models should be audited for how they allocate interpretability, not only for fair predictions.

Auto-discovered on 2026-06-03 by OECD Catalogue Automation

About the metric

Modify this metric

Partnership on AI

Disclaimer: The tools and metrics featured herein are solely those of the originating authors and are not vetted or endorsed by the OECD or its member countries. The Organisation cannot be held responsible for possible issues resulting from the posting of links to third parties' tools and metrics on this catalogue. More on the methodology can be found at https://oecd.ai/catalogue/faq.