GCI: A (G)raph (C)oncept (I)nterpretation Framework

Concept extraction from AI systems now has a measurement tool: GCI provides a quantitative way to evaluate whether concepts discovered by neural networks actually match what humans think those concepts mean—addressing a major gap in explainable AI where extracted concepts were previously difficult to validate objectively.

The framework was tested on molecular property prediction: Researchers demonstrated GCI's practical value by using chemical functional groups (like specific molecular structures) to explain how neural networks make predictions about drug properties, achieving a 0.76 AUCROC completeness score, showing the approach can work on real, complex scientific tasks.

Four concrete applications for responsible AI: GCI enables organizations to (1) evaluate different concept extraction tools, (2) verify that AI-discovered concepts align with human expert understanding, (3) assess whether explanations fully account for how an AI system makes decisions, and (4) apply these methods to high-stakes domains like drug discovery.

Interpretability gaps remain a critical challenge: The paper highlights that even when AI systems extract interpretable concepts, there's often a mismatch between what the AI "learns" and what humans understand—GCI helps measure this gap, which is essential for trustworthy AI deployment in fields like healthcare and chemistry.