Towards Human Cognition Level-based Experiment Design for Counterfactual Explanations (XAI)

One-size-fits-all explanations don't work: Traditional AI explanation systems were designed with technical accuracy in mind rather than considering that different users have different levels of understanding. This paper argues that explanations need to be tailored to match how individual users actually think and learn.

Cognitive science should guide AI transparency: The research proposes using Bloom's taxonomy—an established educational framework for measuring learning levels—to design AI explanations. This means explanations could be adjusted from basic comprehension to more complex analytical thinking depending on the user's cognitive capability, similar to how teachers differentiate instruction for students.

Counterfactual explanations offer a practical path forward: The paper focuses on "counterfactual explanations" (explaining what would need to change for an AI system to make a different decision) as a promising method for helping users understand AI decisions at different cognitive levels, with user feedback used to continuously improve these explanations.

User feedback is critical for validation: Rather than assuming an explanation is good because it's technically correct, this framework emphasizes collecting feedback from actual users at different cognitive levels to verify whether they truly understand the AI's reasoning—and using that feedback to refine future explanations.

Responsible AI requires understanding the user, not just the algorithm: The core insight is that making AI systems truly explainable and trustworthy requires considering human cognition as seriously as we consider the technical aspects of the AI itself.