See What I Mean? CUE: A Cognitive Model of Understanding Explanations

Explanation quality requires more than technical accuracy: Current AI explanation systems often focus on being technically correct rather than cognitively accessible to actual users. The paper's CUE model breaks down understanding into three distinct stages—legibility (can you see it?), readability (can you comprehend it?), and interpretability (can you make sense of it?)—showing that all three matter for genuine human understanding.

Accessibility features don't automatically help everyone: Surprisingly, the study of 455 participants found that colorblind-friendly colormaps like Cividis sometimes worsened outcomes for visually impaired users compared to standard options, challenging the assumption that "accessible" design automatically improves understanding across all disability groups. This suggests one-size-fits-all accessibility solutions are insufficient.

Visually impaired users face real barriers despite equal task performance: While visually impaired participants completed tasks at similar rates as sighted users, they reported lower confidence and required more effort. This gap reveals that current AI explanations may technically work but create unnecessary cognitive burden for people with visual impairments—a hidden equity problem.

AI explanations need to be personalized, not standardized: The findings support moving toward adaptive XAI interfaces tailored to individual users rather than universal explanation formats. Different people may need fundamentally different ways of understanding the same AI decision.

How we present AI decisions affects trust and usability: By demonstrating that changing how explanations are displayed (legibility) directly impacts whether people understand them, the research shows that responsible AI requires attention to presentation design, not just algorithm design.