Behavioral Responses to Simulated Generative-AI Exercise Advice Among University Students: Following, Modifying, Rejecting, and Verifying Recommendations

We examined university students’ stated responses to simulated generative-AI exercise advice. In a randomized 2 × 2 within-participant task, 812 students from four universities in Henan Province evaluated eight fixed, researcher-reviewed scenarios, yielding 6496 decisions. Advice correctness was crossed with the presence of a corresponding reason. Participant-clustered logistic generalized estimating equations modeled direct adoption; a supplementary multinomial model retained following, modifying, rejecting, and verifying as distinct choices. Correct advice was adopted in 85.0% of trials and incorrect advice in 18.8%. Within these packages, reasons increased correct-advice adoption from 79.2% to 90.7% and reduced incorrect-advice adoption from 21.7% to 15.9% (interaction OR = 3.936, 95% CI [3.018, 5.132]). The adjusted adoption gap widened by 17.28 percentage points (95% CI [13.88, 20.67]). Under incorrect advice, the reason-present texts increased rejection, while the change in verification was uncertain. Trust was more strongly associated with adoption under incorrect advice, including after excluding a familiarity item. Associations with the eight-item score were exploratory because of low reliability and prior exposure to matching knowledge questions. The findings distinguish overall adoption from correctness-sensitive choices within standardized exercise-advice scenarios; their relevance to live coaching requires further testing.