Transferable XAI: Relating Understanding Across Domains with Explanation Transfer

Transfer learning can reduce cognitive burden for users encountering related AI systems: Rather than requiring people to independently learn explanations for each new application, Transferable XAI enables users to apply their prior understanding from one domain to another—for example, understanding how BMI affects heart disease risk can help users grasp how it affects diabetes risk, reducing the need for burdensome memorization.

Related AI tasks share explanatory factors but with important differences that matter: The paper demonstrates that while related applications may use the same factors (like BMI in health predictions), these factors can have different relationships to outcomes. The framework accounts for these nuances through an affine transformation approach, helping prevent both dangerous overgeneralization and unnecessary confusion.

User studies confirm Transferable XAI outperforms current explanation approaches: In both formative and summative user studies, participants better understood AI decisions across related domains when using Transferable XAI compared to single-domain explanations or domain-independent approaches, suggesting this method could improve how people interact with multiple AI systems in practice.

The framework addresses a real gap between how AI explanations are currently designed and how people actually use them: Current XAI focuses on explaining individual applications in isolation, but in reality users encounter multiple related AI systems and naturally try to transfer their understanding—this work bridges that gap by designing explanations that explicitly support that transfer process.