Why Don't You Do Something About It? Outlining Connections between AI Explanations and User Actions

Explanations alone don't guarantee action: The paper challenges a core assumption in explainable AI that simply providing explanations will automatically lead users to take meaningful action. This suggests that XAI systems need to be designed with specific user behaviors and outcomes in mind, not just with the goal of making AI systems more transparent.

Current XAI evaluation focuses on the wrong metrics: The research finds that explanations are typically evaluated based on technical aspects rather than whether they actually enable users to do something about AI system decisions. This means many "successful" XAI systems may fail in real-world settings where users need to make decisions or take corrective action.

Significant gaps exist in understanding what information users need: The framework reveals that the XAI community lacks consensus on what types of information explanations should present and which user actions those explanations should support. This fragmentation means there's no clear guidance for building explanations that reliably connect to user decision-making.

The connection between explanation content and user action is underexplored: Prior work "varies widely" in tracing connections between what information is presented and what users actually do with it. This suggests the field needs more empirical research on how different types of explanations influence specific user behaviors in real socio-technical contexts.

Responsible AI requires centering human action from the start: Rather than treating user action as an afterthought, the paper argues that XAI systems should be designed and evaluated based on what actions they enable users to take, making this a critical consideration for developing trustworthy AI systems.