Eye-Tracking Evidence That Verifiable Explanations Support Visual Evidence Checking in AI-Assisted Chest Radiograph...

Evaluations of medical artificial intelligence (AI) explanations often rely on self-reported trust, perceived usefulness, acceptance, or final decision outcomes, while less directly characterizing whether users check evidence around AI outputs during decision making. In AI-assisted chest radiograph interpretation, a critical process-level question is whether users return from the AI output to the original image evidence when further scrutiny is needed. To address this question, we examined whether verifiable explanations-explanations designed to make AI recommendations checkable against the original image evidence-are associated with process markers of visual evidence checking in AI-assisted chest radiograph interpretation using eye-tracking and human-factors process measures. A 2 × 2 between-subjects experiment manipulated verifiable explanations (present vs. absent) and risk context (high vs. low), with AI recommendation correctness embedded at the trial level. Fifty-six clinically trained participants each completed 24 interpretation trials. Analyses focused primarily on gaze transitions between the AI output and the original image and dwell time on the original image, with response time and exploratory verification-related behavioral states used as auxiliary process measures. Verifiable explanations did not simply increase acceptance of AI recommendations. Instead, when AI recommendations were incorrect, they were most clearly associated with more frequent AI-image transitions and longer absolute dwell time on the original image evidence. Exploratory state-based analyses further suggested a lower tendency toward no-verify adopt under incorrect AI recommendations, but these findings were treated as complementary rather than primary evidence. Overall, the value of verifiable explanations lies not only in final decisions but in whether they make AI recommendations more inspectable against the original evidence. These findings provide eye-tracking evidence consistent with visual evidence checking in AI-assisted diagnostic interfaces and underscore the value of process-sensitive human-factors measures in medical AI evaluation.