The Co-12 Recipe for Evaluating Interpretable Part-Prototype Image Classifiers

There is currently no standardized way to evaluate whether explanations from AI image classifiers are actually good. The paper identifies a critical gap in the field: while interpretable part-prototype models (which explain their decisions by showing which visual components they recognized) are gaining popularity, researchers lack agreed-upon criteria for assessing whether these explanations are trustworthy or useful.

The Co-12 framework provides 12 concrete properties for measuring explanation quality, including correctness (are the explanations accurate?), completeness (do they cover all important factors?), and compactness (are they concise enough to understand?). This framework offers a structured approach that the field can use to consistently evaluate and compare different interpretable AI systems.

Part-prototype models offer a promising path toward "explainability by design." Rather than treating explanation as an afterthought, these models are built from the ground up to show their reasoning—they learn recognizable visual parts and explicitly identify which parts led to their classification decision, making them inherently more transparent than black-box alternatives.

The research community needs to establish shared evaluation standards to advance responsible AI development. By outlining research gaps and proposing systematic evaluation approaches, this work emphasizes that responsible AI requires not just building interpretable systems, but also having rigorous methods to verify that the explanations they provide are actually meaningful and reliable.