Confident Doctors Lean Less on AI: Study Links Self-Efficacy to Lower Dependence in Medical Trainees

Generative artificial intelligence has swept into hospitals, laboratories, and lecture halls with astonishing speed, and nowhere is its footprint more visible than in postgraduate medical training. Young physicians racing to publish papers, draft theses, and interpret complex datasets have embraced large language models and AI research assistants at a pace that has outstripped most formal curricula. But a new multicenter study from China raises a provocative question about the psychological profile of those who lean hardest on these tools: are the trainees who depend most on AI also the ones who trust their own abilities least? The answer, according to a survey of more than 1,500 medical postgraduates, appears to be a clear yes.

The research, published in BMC Medical Education, was led by Meng-Pan Li and Meng-Jian Ren of the Department of Orthopedic Oncology at Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine, with colleagues from three institutions across eastern China. Between January and March 2026, the team collected 1,546 valid online questionnaires from medical postgraduates at six participating Chinese universities. Of the respondents, 938, or 60.7 percent, were women, and 579, or 37.5 percent, were studying clinical medicine. The scale and geographic spread of the sample give the findings a weight that smaller, single-campus surveys have lacked, offering one of the most detailed portraits yet of how the next generation of physicians interacts with generative AI in their academic lives.

The study rested on two psychological instruments. Self-efficacy, the well-established construct describing a person’s belief in their own capacity to organize and execute the actions needed to achieve a goal, was measured with a five-item, study-specific academic scale. AI dependence was assessed using a direct Chinese translation of the 22-item Artificial Intelligence Addiction Scale for Researchers, an instrument designed to capture dependence-like patterns of reliance on AI tools rather than simple frequency of use. This distinction matters: the scale probes whether AI use has become compulsive, substitutive, or difficult to regulate, dimensions that echo the structure of behavioral-addiction measures while remaining tailored to the research context of postgraduate trainees.

The statistical architecture of the analysis was deliberately conservative. The primary model was a multivariable linear regression fitted with HC3 heteroscedasticity-robust standard errors, a technique that protects inference when the variability of outcomes differs across subgroups, a common feature of survey data. Restricted cubic splines allowed the team to model the shape of the relationship flexibly rather than forcing it into a straight line. Model diagnostics and, critically, item-overlap sensitivity analyses were performed to guard against the possibility that the two questionnaires shared content that could artificially inflate the association. The researchers also ran an exploratory prediction analysis, using the sample median AI-dependence score of 70 as a non-diagnostic boundary and a fixed 70:30 training-test split for internal validation.

The headline result is striking in its magnitude and precision. In the primary adjusted model, each 5-point increase in self-efficacy score was associated with a 5.296-point decrease in AI-dependence score, with a 95 percent confidence interval spanning from minus 5.918 to minus 4.674 and a P value below 0.001. Restricted cubic splines confirmed the overall association, with a chi-square of 287.10 on three degrees of freedom, and revealed modest but statistically detectable nonlinearity, with a chi-square of 8.79 on two degrees of freedom and a P value of 0.012. In other words, the inverse relationship is broadly consistent across the range of self-efficacy scores, though the slope may bend slightly at the extremes.

Skeptics of questionnaire-based research often point to a subtle threat: when two instruments contain items that measure similar content, the apparent association between their scores can be an artifact of wording rather than a genuine relationship between constructs. The authors confronted this problem head-on. After removing clearly overlapping items from both the self-efficacy and AI-dependence scales, the estimate from the fully adjusted Model 3 changed by only 2.1 percent, settling at a beta of minus 5.187 with a 95 percent confidence interval of minus 5.800 to minus 4.575. Supplementary analyses reconstructed the scores with individual items excluded, rescaling them to their original ranges, and the inverse association held throughout. The signal, in short, is not a measurement artifact.

The exploratory prediction component added a machine-learning dimension to the epidemiological one. Using the sample median of 70 as a cut point, the team built a model to identify trainees with sample-relative higher AI-dependence scores. In the held-out internal test set, the model achieved an area under the receiver operating characteristic curve of 0.875, with a 95 percent confidence interval of 0.845 to 0.906. An AUC near 0.88 indicates good discriminative ability, meaning the model could meaningfully separate higher-dependence from lower-dependence trainees within this sample. The authors are careful, however, to label this boundary as sample-relative rather than diagnostic: a score of 70 has no established clinical meaning, and the model was validated internally only, on data drawn from the same cohort.

What might explain the inverse relationship between confidence in one’s own abilities and dependence on AI? The authors suggest that self-efficacy may act as a psychological buffer: trainees who trust their capacity to plan, execute, and verify academic tasks may treat AI as an optional accelerator rather than a cognitive crutch. Conversely, those who doubt their own competence may reach for generative tools to fill perceived gaps, and repeated outsourcing of effort can, in a feedback loop, further erode independent skill. The cross-sectional design of the study, capturing a single moment in time, cannot determine whether low self-efficacy drives AI dependence, whether heavy AI use undermines self-efficacy, or whether both are shaped by a third factor such as academic workload, anxiety, or institutional culture. The authors explicitly state that causality cannot be inferred.

The practical implications, nonetheless, are pointed. The researchers argue that AI-literacy education in postgraduate medical programs may need to go beyond teaching how to prompt a model or spot a hallucinated citation. Their findings suggest such curricula should also support autonomous task engagement and critical verification of AI outputs, cultivating the self-directed confidence that appears to protect against dependence-like use. If a resident believes she can design the analysis herself, the argument runs, she is more likely to use AI as a checkable assistant and less likely to accept its output uncritically. That framing aligns with a growing international conversation about academic integrity, cognitive offloading, and the skills that physicians of the AI era must retain regardless of how capable the tools become.

The study also carries important caveats that temper any temptation toward alarm. The sample was drawn from six Chinese universities, and whether the self-efficacy-AI dependence gradient generalizes to trainees in other educational systems remains to be tested. The self-efficacy measure was study-specific rather than a validated general instrument, and the AI-dependence scale, while adapted from an established researcher-focused tool, was a direct translation. The prediction model requires independent validation before any practical use, and the median-based cut point is a statistical convenience, not a diagnostic threshold. Funded by the National Natural Science Foundation of China and other national programs, and approved by the ethics committee of Shanghai General Hospital with electronic informed consent from all participants, the study was conducted under the Declaration of Helsinki. As generative AI becomes as routine in medical training as the stethoscope once was, this research offers a rare quantitative anchor: the strongest safeguard against over-reliance on machines may be a well-calibrated belief in oneself.

Subject of Research: The association between academic self-efficacy and AI dependence among medical postgraduates

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

Article References: Association between self-efficacy and AI dependence among medical postgraduates: a multicenter cross-sectional study with exploratory prediction. (n.d.). https://doi.org/10.1186/s12909-026-10539-6

Image Credits: AI Generated

DOI: 10.1186/s12909-026-10539-6

Keywords: artificial intelligence, self-efficacy, medical education, postgraduate training, AI dependence, cross-sectional study, machine learning, psychometrics, generative AI, medical postgraduates, survey research, China