Association between self-efficacy and AI dependence among medical postgraduates: a multicenter cross-sectional study with exploratory prediction

Generative artificial intelligence (AI) is increasingly used in postgraduate medical education and research, but the relationship between self-efficacy and dependence-like reliance on AI remains insufficiently characterized. We examined the association between continuous self-efficacy and AI-dependence scores and separately explored prediction of sample-relative higher AI-dependence scores. This multicenter cross-sectional online survey included 1,546 valid questionnaires from medical postgraduates at six participating Chinese universities between January and March 2026. Self-efficacy was assessed using a five-item study-specific academic measure, and AI dependence was assessed using a direct Chinese translation of the 22-item Artificial Intelligence Addiction Scale for Researchers. Analyses used multivariable linear regression with HC3 standard errors, restricted cubic splines, model diagnostics, and item-overlap sensitivity analyses. Exploratory prediction used the sample median score of 70 as a non-diagnostic boundary and a fixed 70:30 training/test split for internal validation. Among 1,546 participants, 938 (60.7%) were women and 579 (37.5%) studied clinical medicine. In the primary adjusted model, each 5-point higher self-efficacy score was associated with a 5.296-point lower AI-dependence score (95% CI, -5.918 to -4.674; P < 0.001). Restricted cubic splines supported the overall association (chi-square = 287.10; df = 3; P < 0.001) and modest nonlinearity (chi-square = 8.79; df = 2; P = 0.012). After removal of clearly overlapping items from both instruments, the Model 3 estimate changed by 2.1% (beta = -5.187; 95% CI, -5.800 to -4.575). The exploratory model achieved an AUC of 0.875 (95% CI, 0.845–0.906) in the held-out internal test set. Higher self-efficacy was inversely associated with continuous AI-dependence scores among medical postgraduates. These findings suggest that AI-literacy education may need to support autonomous task engagement and critical verification of AI outputs, but the cross-sectional design precludes causal interpretation. The exploratory prediction analysis was internally validated only, used a sample-relative boundary, and requires independent validation before practical use.