When recommendations are explainable An eye-tracking study comparing how and what to explain

This study investigates how explanation scope affects users’ understanding, trust, and attitude toward a recommender system (RS), and how these effects are mediated by visual attention and moderated by explanation modality. Drawing on eye-tracking data, we conducted a 3 (explanation scope: local, global, joint) × 2 (modality: text-only, text with illustrations) between-subjects experiment, with a control group with no explanation (N = 277). Results showed that joint explanations outperformed no-explanation conditions across all outcomes. Structural equation modeling revealed that visual attention mediated the positive effects of joint explanations on users’ understanding of the RS. However, graphic illustrations can undermine the benefits of joint explanations, particularly in terms of users’ evaluation of the RS. These findings suggest that explanation effectiveness depends on attentional engagement and design alignment, offering a human-centered framework for designing adaptive, cognitively efficient explanations that support meaningful user interaction with AI systems.