Presentation of AI-based decision support influences clinician decision making

Abstract Background Artificial intelligence (AI)-based clinical decision support is increasingly being incorporated into dental diagnosis. While AI performance has been extensively studied, less is known about how different presentations of AI recommendations influence clinicians’ diagnostic decisions. Methods This randomized experimental study evaluated the effects of AI presentation format on radiographic furcation detection among dental faculty and predoctoral students. Participants were randomly assigned to one of three conditions: (1) no AI assistance (control), (2) visual AI annotation, or (3) descriptive (text-based) AI recommendation. The prespecified primary outcome was overall diagnostic accuracy across the five radiographic cases, compared among the three randomized presentation groups. Secondary outcomes included diagnostic confidence, interpretation time, and perceived usefulness of AI. Results A total of 122 participants (60 faculty and 62 students) completed the study: 41 were assigned to control, 40 to visual AI annotation, and 41 to descriptive AI recommendation. In the prespecified primary analysis, pooled diagnostic accuracy across the five radiographic cases was 87.8% in the control group, 84.0% in the visual AI group, and 84.4% in the descriptive AI group, with no significant difference among presentation formats (χ² = 1.434; p = 0.488). Secondary exploratory analyses identified a significant presentation effect for tooth #18 among students, for whom visual AI annotation was associated with greater agreement with an incorrect AI recommendation. Faculty receiving visual AI output demonstrated reduced post-task confidence compared with students, whereas perceptions of AI usefulness remained generally positive across groups. Interpretation time differed between faculty and students for selected radiographs. Conclusions AI presentation format may influence clinician responses during complex diagnostic scenarios, particularly when radiographic interpretation is challenging. However, these effects were not observed consistently across all cases, suggesting that the influence of AI presentation may be context-dependent and warrants further investigation.