This cross-sectional survey was conducted between January and February 2026 at the University of Ulsan College of Medicine to examine factors associated with AI use intention among medical students and faculty members. Specifically, this study compared AI-related characteristics, AI perceptions, AI literacy, and AI use intention between the 2 groups and identified factors independently associated with AI use intention within each group.

Eligible participants included all enrolled second-year premedical students and first- through fourth-year medical students (n=200), as well as all full-time faculty members (n=532) at the University of Ulsan College of Medicine. First-year premedical students were excluded because the survey was conducted before their enrollment.

An invitation containing information about the study and a link to the anonymous web-based questionnaire was distributed through the institutional communication system. Participation was voluntary, and no incentives were provided.

A total of 235 participants completed the survey, including 141 (60%) medical students (response rate: 141/200, 70.5%) and 94 (40%) faculty members (response rate: 94/532, 17.7%). Only complete questionnaires were included in the final analyses. This online survey is reported in accordance with the CHERRIES (Checklist for Reporting Results of Internet E-Surveys) guidelines (Checklist 1).

This study was approved by the Institutional Review Board of Asan Medical Center (S2025-2554-0001). Before accessing the questionnaire, all participants received information regarding the study purpose, voluntary participation, confidentiality, and data protection. Electronic informed consent was obtained before participation.

The survey was conducted anonymously, and no personally identifiable information was collected. All data were stored in deidentified form and used exclusively for research purposes.

Statistical analyses were performed using Jamovi (version 2.7.24). Continuous variables are presented as means and SDs, whereas categorical variables are presented as frequencies and percentages. The internal consistency of all multi-item scales was evaluated using Cronbach α coefficients, with values greater than 0.70 considered acceptable. Skewness and kurtosis were examined to assess the assumption of univariate normality.

Baseline characteristics were compared between medical students and faculty members using independent-sample 2-tailed t tests for continuous variables and chi-square tests for categorical variables. Homogeneity of variances was assessed using the Levene test, and the Welch t test was applied when the assumption of equal variances was violated.

Unadjusted between-group comparisons were considered the primary descriptive comparisons. Exploratory analyses of covariance additionally adjusted for sex, age, and previous AI education because these characteristics differed between groups. However, age was strongly associated with learner group, and the age distributions of students and faculty had limited overlap. Therefore, age-adjusted group estimates were treated as sensitivity analyses and were not interpreted as representing a causal or directly comparable student-faculty contrast.

Pearson correlation analyses were conducted separately for medical students and faculty members. Group-specific multiple linear regression models were estimated to describe associations with AI use intention. To directly test whether predictor associations differed by participant group, a pooled regression model included participant group and group-by-predictor interaction terms for perceived benefits, perceived risks, total AI literacy, and AI use frequency. Continuous predictors were mean centered before interaction terms were constructed. The pooled model additionally included AI knowledge, AI familiarity, sex, age, and previous AI education. Because AI use intention was measured on a bounded 5-point scale and showed a marked ceiling effect, HC3 heteroscedasticity–robust SEs were used for primary inferential interpretation. Ordinal logistic regression was conducted as a sensitivity analysis using the same predictor set. For this analysis, perceived benefits, perceived risks, total AI literacy, and AI use frequency were standardized before constructing the group-by-predictor interaction terms; thus, the corresponding odds ratios represent associations per 1-SD increase. The proportional odds assumption was also assessed. Complete ordinal logistic regression results, including main effects, interaction terms, threshold estimates, model fit statistics, and assessment of the proportional odds assumption, are provided in Table S2 in Multimedia Appendix 1. Additional sensitivity analyses examined reduced predictor sets to evaluate the stability of the student AI literacy coefficient. Multicollinearity was assessed using variance inflation factors and tolerance statistics. Results are reported as unstandardized coefficients; standardized coefficients where applicable; odds ratios for ordinal logistic regression; 95% CIs; and 2-sided P values, with P<.05 considered statistically significant.

A total of 235 participants completed the survey, comprising 141 (60%) medical students and 94 (40%) faculty members (Table 1). Among the students, 58.2% (82/141) were male, and the mean age was 23.6 (SD 1.87) years. Second-year premedical students constituted the largest academic year subgroup (38/141, 27.0%). Regarding intended career paths, surgery was the most frequently selected option (53/141, 37.6%).

Among the faculty members, 71.3% (67/94) were male, and the mean age was 48.9 (SD 7.58) years. Half (47/94, 50%) of the faculty participants were professors. Regarding medical specialties, other clinical specialties represented the largest category (45/94, 47.9%). Their mean professional experience was 17.3 (SD 8.19) years.

In the unadjusted comparisons, significant between-group differences were observed for sex (χ²₁=4.2; P=.04), age (t_100.6=−31.7; P<.001), and previous AI education (χ²₁=18.2; P<.001; Table 2). Faculty members included a higher proportion of male individuals, whereas students reported greater previous AI education experience. Because age was strongly confounded with learner group and the observed age ranges had limited overlap, the covariate-adjusted estimates that follow are presented as exploratory sensitivity analyses and should not be interpreted as comparisons of otherwise equivalent students and faculty members.

In the exploratory covariate-adjusted analysis, faculty members reported more frequent AI use than students (adjusted mean 4.13, SE 0.224 vs 3.30, SE 0.159; F_1,230=5.550; P=.02). No adjusted differences were observed in AI knowledge level (F_1,230=0.132; P=.72) or AI familiarity (F_1,230=0.304; P=.58). These adjusted findings should be interpreted cautiously because of the limited age overlap between groups.

The AI perception and AI literacy instruments demonstrated acceptable internal consistency in the present sample. The Cronbach α was 0.884 for the perceived benefits subscale and 0.617 for the perceived risks subscale. For the AI literacy instrument, Cronbach α values ranged from 0.726 to 0.847 across the 4 domains, and the overall scale demonstrated excellent internal consistency (Cronbach α=0.899).

In exploratory analysis of covariance sensitivity analyses adjusting for sex, age, and previous AI education, no statistically significant differences were observed between medical students and faculty members in perceived benefits, perceived risks, AI use intention, or any AI literacy domain (P>.05 in all cases; Table 3). Given the strong association between age and learner group and limited age overlap, these adjusted means are descriptive sensitivity estimates rather than evidence of equivalent student-faculty populations. Item-level analysis of the perceived-risk scale in the full sample showed mean scores of 3.65 (SD 1.10), 3.81 (SD 0.91), and 3.93 (SD 0.94) for impaired clinical judgment, potential diagnostic errors, and privacy concerns, respectively; corrected item-total correlations ranged from 0.391 to 0.455.; detailed item-level results are provided in Table S1 in Multimedia Appendix 1.

Correlation analyses were conducted separately for medical students and faculty members (Tables 4 and 5). Among medical students, perceived benefits showed a strong positive correlation with AI use intention (r=0.660; P<.001), whereas total AI literacy was weakly and nonsignificantly associated with AI use intention (r=0.155; P=.07). Perceived risks demonstrated a weak positive correlation with AI use intention (r=0.293; P<.001).

Among faculty members, both perceived benefits (r=0.588; P<.001) and total AI literacy (r=0.408; P<.001) were positively correlated with AI use intention, whereas perceived risks were not significantly associated with AI use intention (r=0.140; P=.18).

Group-specific multiple linear regression models retaining the full covariate set (perceived benefits, perceived risks, AI literacy, AI use frequency, AI knowledge, AI familiarity, sex, age, and previous AI education) were estimated for students and faculty members (Table 6). HC3 heteroscedasticity–robust SEs were used as the primary basis for inference because the bounded intention outcome showed a marked ceiling effect.

Among faculty members, perceived benefits were positively associated with AI use intention (B=0.529, HC3 95% CI 0.288-0.771; standardized β=0.503; P<.001). AI use frequency also remained positively associated with intention (B=0.137, HC3 95% CI 0.025-0.249; standardized β=0.257; P=.02). AI literacy was positive in the conventional linear model but marginal with HC3 inference (B=0.221, HC3 95% CI −0.018 to 0.460; P=.07).

In the pooled HC3-robust regression model (N=235), the group-by–AI literacy interaction was statistically significant (B=0.361, 95% CI 0.083-0.638; P=.01), indicating that the adjusted association between AI literacy and use intention differed between students and faculty members. The interactions for perceived benefits, perceived risks, and AI use frequency were not significant (Table 7). The model explained 47.0% of the variance (adjusted R²=0.436; F_14,220=11.88; P<.001).

AI use intention showed a marked ceiling effect: 55.3% (130/235) of participants selected “strongly agree” (score 5), and 36.6% (86/235) selected “agree” (score 4). In ordinal logistic sensitivity analyses, the group-by–AI literacy interaction remained significant (odds ratio 2.48, 95% CI 1.24-4.98; P=.01), whereas the other interactions remained nonsignificant. Complete ordinal logistic regression results, including main effects, interaction terms, threshold estimates, model fit statistics, and assessment of the proportional odds assumption, are provided in Table S2 in Multimedia Appendix 1. The student AI literacy coefficient was not stable across reduced linear models, and the positive student perceived risk coefficient was not significant with HC3 robust inference. These findings support limiting the principal group difference claim to the AI literacy interaction and interpreting subgroup coefficients cautiously.

This study examined factors associated with broad self-reported AI use intention among medical students and faculty members within a single medical school. Three principal findings emerged.

First, perceived benefits were consistently and strongly associated with AI use intention in both learner groups. This association remained robust across the primary regression analyses and sensitivity analyses using HC3 heteroscedasticity–robust SEs. These findings suggest that participants who perceived AI as beneficial for education, research, or clinical practice were more likely to report intentions to use AI regardless of learner group.

Second, the pooled interaction analysis demonstrated that the adjusted association between AI literacy and AI use intention differed significantly between medical students and faculty members. Importantly, this conclusion was based on direct statistical testing of the interaction rather than comparison of separate subgroup regressions. However, the subgroup-specific pattern should be interpreted cautiously. Among students, the zero-order association between AI literacy and use intention was weak and positive, whereas the adjusted regression coefficient became negative only after controlling for correlated AI-related variables and varied across reduced models. This pattern is compatible with suppression or shared variance rather than a stable inverse relationship. In contrast, among faculty members, AI literacy showed a positive association with AI use intention that was generally consistent across analyses.

Third, although perceived risks appeared positively associated with AI use intention among students in the conventional regression model, this association was attenuated and no longer statistically significant when HC3-robust SEs were applied. Moreover, the pooled interaction model did not demonstrate a significant difference between students and faculty members for perceived risks. Accordingly, this finding should be regarded as exploratory rather than evidence that greater awareness of AI-related risks increases willingness to use AI.

The finding that perceived benefits represented the strongest correlate of AI use intention is consistent with those of previous studies examining technology acceptance in health care. The TAM and UTAUT consistently identify perceived usefulness or performance expectancy as the principal determinant of behavioral intention. Similarly, recent studies involving physicians, medical students, and other health care professionals have reported that expected educational or clinical benefits are the most influential predictors of AI acceptance. Our findings extend this literature by demonstrating that perceived benefits remained the dominant correlate of AI use intention even after accounting for AI literacy, perceived risks, AI familiarity, AI knowledge, and previous AI education.

The relationship between AI literacy and AI use intention was more complex. Previous studies have generally assumed that greater AI literacy facilitates AI adoption because users become more confident in understanding and applying AI technologies. However, there is emerging evidence also suggesting that greater AI literacy may increase awareness of AI limitations, uncertainty, ethical concerns, and the need for critical evaluation. Rather than demonstrating that AI literacy has uniformly positive or negative effects, our findings suggest that its association with AI use intention may depend on learners’ educational and professional contexts. Faculty members who routinely integrate AI into teaching, research, or clinical work may experience AI literacy primarily as a practical competency, whereas students may simultaneously acquire greater awareness of both the opportunities and limitations of AI during training. Because the adjusted student coefficient was sensitive to model specification, these findings should be interpreted as evidence of contextual variation rather than evidence that greater AI literacy reduces students’ willingness to use AI.

Our findings also highlight the importance of distinguishing between broad AI use and specific AI applications. The present questionnaire assessed general intention to use AI across education, research, and clinical practice rather than focusing on a particular AI technology such as generative AI or clinical decision support systems. These contexts differ substantially with respect to task complexity, responsibility, explainability requirements, governance, and acceptable levels of risk. Recent studies suggest that trust, governance, institutional policies, and clinical context substantially influence AI adoption in health care. For example, a 2026 survey of health care practitioners found that trust and perceived risk jointly shaped AI acceptance and readiness, with trust functioning as a construct distinct from perceived benefit or risk alone [23], and work on health chatbot adoption has shown that perceived risk lowers trust both directly and indirectly through perceived usefulness and self-efficacy [24]. Separately, studies of generative AI integration into medical curricula indicate that adoption is shaped substantially by institutional governance, including policy on permitted use and equitable platform access, in addition to individual perceptions of benefit or risk [25]. Because trust and governance constructs were not directly measured in the present study, future investigations should incorporate them explicitly, distinguish among different categories of AI technologies, and examine how determinants of AI acceptance vary across educational and clinical settings.

Although the present findings are exploratory, they suggest several implications for AI education in medical schools. The significant interaction between learner group and AI literacy indicates that identical educational approaches may not address the needs of students and faculty members equally. Future educational interventions for students may benefit from emphasizing supervised AI application, critical appraisal of AI-generated information, appropriate calibration of trust, and responsible clinical reasoning rather than focusing exclusively on conceptual knowledge. For faculty members, educational programs may place greater emphasis on integrating AI into authentic teaching, research, and clinical activities while supporting confident and appropriate implementation.

Importantly, these implications should be regarded as educational hypotheses rather than evidence-based curricular recommendations. Because this study did not evaluate educational interventions, curriculum design, or subsequent AI use behavior, further longitudinal and intervention studies are needed before specific educational strategies can be recommended.

Several limitations should be acknowledged.

First, participants were recruited from a single medical school, and the faculty response rate was relatively low. Consequently, self-selection bias cannot be excluded, and the findings may not be generalizable to other institutions or health care settings.

Second, the cross-sectional design and reliance on self-reported questionnaire data preclude causal inference. The observed associations cannot determine whether AI use frequency increases AI use intention, whether stronger intentions lead to more frequent AI use, or whether both are influenced by other unmeasured factors.

Third, AI use intention was assessed using a single broad item encompassing education, research, and clinical practice. Although this approach captured participants’ overall intention to use AI, it could not distinguish among different AI technologies or application contexts and did not permit assessment of measurement error. Future studies should use validated multi-item measures and separately examine intentions related to generative AI, clinical decision support systems, and other AI applications.

Fourth, although the AI literacy instrument was adapted from established theoretical frameworks and demonstrated acceptable internal consistency, additional psychometric evaluation—including exploratory or confirmatory factor analysis, construct validation, and measurement invariance across learner groups—was beyond the scope of the present study. Future research should further validate the measurement properties of this instrument in larger and more diverse samples.

Finally, the perceived risk subscale demonstrated relatively limited internal consistency and combined concerns related to clinical judgment, diagnostic errors, and privacy. Furthermore, the positive association between perceived risks and AI use intention among students was not robust in HC3 sensitivity analyses, and the corresponding interaction effect was not statistically significant. Consequently, findings related to perceived risks should be interpreted cautiously until replicated in future studies using more refined measures of AI-related risk.

Perceived benefits were consistently associated with stronger broad self-reported AI use intention among both medical students and faculty members. The association between AI literacy and AI use intention differed according to learner group, although the underlying mechanism remains uncertain. These findings provide exploratory evidence that learner context may shape AI use intention and highlight the need for multicenter longitudinal studies using validated measures to better understand AI adoption in medical education.