ABSTRACT

Objectives

Resident physicians increasingly encounter artificial intelligence (AI), but perceived readiness may reflect attitudes and other self-assessments as well as exposure to AI. This exploratory study examined self-reported medical AI readiness in relation to AI attitudes, AI anxiety, interprofessional identity, and prior AI- or health-technology-related training.

Materials and Methods

In this single-center cross-sectional online survey, all 136 residents in psychiatry, child and adolescent psychiatry, pediatrics, and family medicine at a university-affiliated training hospital in Türkiye were invited; 84 participated (61.8%). Associations among scale totals were examined using Spearman correlations with bootstrapped 95% confidence intervals (CIs) and Benjamini-Hochberg correction. A theory-informed post-hoc regression model and sensitivity analyses were used to examine independent associations with readiness.

Results

Mean scores were 64.77 [standard deviation (SD) 16.40] for medical AI readiness, 112.74 (SD 17.67) for AI attitudes, 45.90 (SD 8.71) for interprofessional identity, and 70.43 (SD 27.63) for AI anxiety. AI attitudes correlated with readiness (ρ=0.32, 95% CI 0.11-0.50, q=0.018), interprofessional identity (ρ=0.26, q=0.038), and AI anxiety (ρ=-0.29, q=0.020). In the regression model, more positive attitudes were associated with higher readiness (b=0.33, standardized β=0.36, 95% CI 0.11-0.55, p=0.004); estimates for anxiety, interprofessional identity, and prior training were imprecise. The attitude coefficient remained positive in sensitivity analyses examining influential observations, robust standard errors, and adjustment for demographic characteristics and specialty.

Conclusion

More positive AI attitudes were associated with higher self-reported medical AI readiness in this single-center sample. Partial construct overlap and common-method variance may contribute to this association. The findings do not establish objective competence or causality; prospective studies using observed performance should evaluate clinically grounded AI education.

Introduction

Artificial intelligence (AI) is increasingly used in healthcare for diagnostic classification, risk prediction, monitoring, clinical decision support, and education. In psychiatry and mental healthcare, AI-supported approaches have been investigated across diagnosis, monitoring, and intervention, while questions remain about clinical utility, patient safety, governance, and ethical implementation.^{1, 2} Medical education faces a parallel challenge: clinicians must understand not only how AI systems may support practice, but also their limitations, data requirements, potential biases, and appropriate role in human decision-making.^{3, 4}

Preparedness for medical AI is broader than frequency of technology use. The Medical Artificial Intelligence Readiness Scale for Medical Students (MAIRS-MS) conceptualizes perceived readiness across cognition, ability, vision, and ethics.⁵ Validation studies in several languages have supported this multidimensional structure,^{6, 7} while surveys of medical and health-professions students commonly report substantial interest alongside limited formal instruction and varying levels of perceived preparedness.^{8, 9} Studies of postgraduate trainees similarly report educational gaps and concerns about responsible clinical integration.^10-12

In the conceptual model guiding this exploratory study, AI attitudes, AI anxiety, interprofessional identity, and previous training were treated as distinct candidate correlates of perceived readiness. Favorable attitudes may reflect willingness to learn, perceived usefulness, and confidence; anxiety may reflect learning demands, job displacement, sociotechnical risks, and human-like AI configurations.^13-16 Previous training may provide exposure to relevant concepts, although heterogeneous training cannot be assumed to represent a standardized curriculum. The present study tested only direct associations. It was not designed or powered to evaluate mediation, moderation, or indirect pathways.

Interprofessional identity was included because clinical AI is implemented within team-based care rather than by isolated users. Extended professional identity theory describes belonging, commitment, and beliefs concerning collaborative practice.^17-20 AI-supported decisions may redistribute information, roles, accountability, and communication across clinical teams. Interprofessional identity could therefore relate to openness toward new team-mediated technologies, although it is not itself a measure of technological competence. This rationale remains tentative in a sample consisting exclusively of physicians, and direct evidence remains limited. Consistent with this, an international study of mental health practitioners emphasized psychological, ethical, and professional factors in intentions to adopt AI-enabled technologies.²¹

The four departments were selected a priori because they had established clinical and educational collaboration at the study center and collectively represented mental healthcare, child-focused care, and longitudinal generalist care. These contexts rely heavily on communication, shared decision-making, and coordination across professional boundaries, while emerging postgraduate literature highlights AI-related educational needs in pediatrics and across resident groups.^{12, 22} The selection was not intended to represent all residency programs or technology-intensive specialties.

The primary objective was to examine whether AI attitudes, AI anxiety, interprofessional identity, and broadly defined prior AI- or health-technology-related training were independently associated with self-reported medical AI readiness. Secondary objectives were to describe self-reported AI exposure and training, examine pairwise relationships among the four constructs, and evaluate consistency between MAIRS-MS and a single global self-rating of clinical AI preparedness. Because the study had no preregistered hypotheses or confirmatory analysis plan, all inferential analyses are described as exploratory.

Materials and Methods

Design and Setting

This single-center cross-sectional online survey was conducted from April 1 to May 1, 2026, at Bolu Abant İzzet Baysal University-affiliated training and research hospital in Türkiye. The survey constituted the quantitative component of the project. An optional qualitative follow-up had been considered when the original survey form was developed, but it was not undertaken; no qualitative data were collected or analyzed.

Participants, Sampling, and Survey Administration

Eligibility required active residency in psychiatry, child and adolescent psychiatry, pediatrics, or family medicine at the study institution and voluntary electronic informed consent. No additional inclusion or exclusion criteria were applied. Other residency programs existed at the center but were outside the a priori sampling frame. The four included departments were selected because of established clinical and educational collaboration and their shared involvement in mental health, child-focused, and longitudinal generalist care. All 136 eligible residents in the selected departments were invited: 18 in child and adolescent psychiatry, 27 in pediatrics, 73 in family medicine, and 18 in psychiatry.

Each eligible resident received an individually generated Google Forms survey link, distributed once by the research team. Google account-based “limit to 1 response” functionality was enabled. Google account identifiers and IP addresses were not collected, stored, or visible in the research dataset. Electronic informed consent was mandatory before the form could advance. All 84 respondents provided electronic informed consent and complete scale data and were included in the analysis. No respondent was excluded because of incomplete responses or for any other reason; no imputation was required. The overall response rate was 61.8%: child and adolescent psychiatry, 18/18 (100%); pediatrics, 17/27 (63.0%); family medicine, 36/73 (49.3%); and psychiatry, 13/18 (72.2%).

A yes/no item asked about willingness to participate in a possible qualitative follow-up. Contact information, when volunteered, was collected through a separate form and was not linked to the survey dataset. The qualitative follow-up was not conducted and the item was not analyzed.

Sociodemographic and AI-Exposure Form

The form recorded age, self-reported sex, marital status, childhood residential setting, medical school, graduation year, residency start year, department, residency year, daily professional digital-device use, the frequency and purposes of using tools that participants perceived as AI-based, prior AI- or health-technology-related training, and a single global preparedness item. Professional device use was recorded in five categories (<1, 1-2, 3-4, 5-6, or ≥7 hours/day). AI-use frequency ranged from never to almost daily. Participants could select more than one purpose of use: medical information/clinical queries, academic writing/literature searching, patient-education materials, clinical decision support, and personal/everyday use.

The questionnaire provided examples of general-purpose tools (ChatGPT, Gemini, and Copilot) but did not record application names or define “clinical decision support.” Consequently, all exposure findings represent participant-interpreted purposes rather than verified use of regulated clinical systems or use during patient care. The training item asked whether participants had received “AI/health-technology” education in medical school, residency, both, or through an independent online course. No definition was provided for qualifying training; no participant selected the independent-course option. The global item asked, “How ready do you generally feel to use AI in clinical practice?” with responses ranging from 1 (not at all ready) to 5 (fully ready).

Measures

Medical Artificial Intelligence Readiness Scale for Medical Students (MAIRS-MS)

The 22-item MAIRS-MS assesses perceived readiness across cognition (items 1-8), ability (items 9-16), vision (items 17-19), and ethics (items 20-22), using a 5-point response format. Total scores range from 22 to 110, with higher scores indicating greater perceived readiness.⁵ The validated Turkish wording was administered without modification. Although developed for medical students, it was selected as a structured multidimensional self-report measure; its psychometric properties in residents remain to be established.

Artificial Intelligence Attitude Scale

This 29-item Turkish instrument assesses interest in learning about AI, reluctance to adopt AI, belief in AI capabilities, and resistance to AI.²³ Items are rated from 1 to 5. Items 10-16 and 24-29 are reverse-scored so that higher total scores indicate more positive attitudes. Possible total scores range from 29 to 145. Several items concern interest, confidence, or favorable self-appraisal; these domains are conceptually distinct from medical AI readiness but may partially overlap with MAIRS-MS self-perceptions.

Extended Professional Identity Scale-Turkish (EPIS-TR)

The 12-item EPIS-TR assesses interprofessional belonging, commitment, and beliefs using a 5-point response format. No items are reverse-scored; total scores range from 12 to 60, with higher scores indicating stronger interprofessional identity.^{18, 19} It was validated in multidisciplinary health-professions samples rather than an exclusively physician resident sample.

Artificial Intelligence Anxiety Scale

The Turkish 21-item scale uses a 7-point response format and assesses learning anxiety (items 1-8), job-replacement anxiety (items 9-14), sociotechnical-blindness anxiety (items 15-18), and AI-configuration anxiety (items 19-21). Total scores range from 21 to 147, with higher scores indicating greater anxiety.^{13, 24}

Statistical Analysis

Scale scores were calculated according to published instructions. Internal consistency was assessed using Cronbach’s α and McDonald’s ω total for the full scales and subscales after reverse scoring. These coefficients were treated as reliability estimates, not evidence of construct validity. Continuous variables are summarized as means [standard deviations (SDs)] and medians (interquartile ranges), and categorical variables with counts and percentages.

Pairwise associations among the four scale totals were evaluated using Spearman correlations. Bootstrap 95% confidence intervals (CIs) were estimated using 5,000 resamples. Benjamini-Hochberg false discovery rate (FDR) correction was applied across the six pairwise correlations.²⁵ The global preparedness item was analyzed separately because of semantic overlap with MAIRS-MS. Exploratory correlations between four AI-attitude and four MAIRS-MS subscales formed a separate 16-test family with Benjamini-Hochberg correction.

A theory-informed post-hoc multiple linear regression used MAIRS-MS total score as the dependent variable and AI-attitude total, AI-anxiety total, EPIS-TR total, and any previous AI/health-technology training (0=no; 1=any) as predictors. The global preparedness item was excluded because of conceptual overlap. The intercept, unstandardized and standardized coefficients, 95% CIs, and model-fit statistics are reported. Coefficient-level q-values were calculated with Benjamini-Hochberg correction across the four focal predictors.

Diagnostic assessments included residual and quantile-quantile (Q-Q) plots, the Shapiro-Wilk test, the Breusch-Pagan test, variance inflation factors (VIFs), standardized residuals, leverage, Cook’s distance, and exploratory quadratic terms for continuous predictors. Sensitivity analyses used type 3 heteroscedasticity-consistent (HC3) robust standard errors, removed the observation with the largest Cook distance, adjusted for age, sex, and residency year, and added specialty as a categorical fixed covariate (with child and adolescent psychiatry as the reference category). Specialty and subscale analyses were exploratory and not used to generate confirmatory claims.

All eligible residents in the selected departments were invited; recruitment was not based on an a priori power calculation. A sensitivity calculation for the achieved N=84 indicated 80% power at two-sided α=0.05 to detect approximately |ρ|=0.30 and, for a four-predictor regression, Cohen f²=0.15. Smaller associations and subgroup effects could therefore remain undetected. Retrospective observed power was not calculated. Analyses were conducted in IBM SPSS Statistics 26.0 (IBM Corp., Armonk, NY, USA). The bootstrap seed was 20260805. Reporting followed STROBE and CHERRIES guidance.^{26, 27}

Ethics

The Bolu Abant İzzet Baysal University Non-Interventional Clinical Research Ethics Committee approved the study (approval no: 2026/120; date: March 24, 2026). The study followed the Declaration of Helsinki. Electronic informed consent was obtained before survey access. Google account identifiers, IP addresses, and direct identifiers were not retained in the analytic dataset.

Results

Participant Characteristics, Response, and AI Exposure

Of 136 invited residents, 84 participated (61.8%). Response rates varied by department: 100% in child and adolescent psychiatry, 72.2% in psychiatry, 63.0% in pediatrics, and 49.3% in family medicine. Mean age was 28.30 years (SD 2.42; range 25-37); 52 participants (61.9%) were female. Family medicine accounted for 42.9% of respondents, followed by child and adolescent psychiatry (21.4%), pediatrics (20.2%), and psychiatry (15.5%).

One participant (1.2%) reported never using tools perceived as AI-based, 7 (8.3%) use a few times per month, 31 (36.9%) use several times per week, and 45 (53.6%) almost-daily use. The most frequent self-reported purposes were personal/everyday tasks (91.7%), medical information/clinical queries (85.7%), and academic writing/literature searching (64.3%). Sixty-nine participants (82.1%) reported no prior training; 2 reported training in medical school, 11 during residency, and 2 in both settings. These categories do not indicate training content or effectiveness (Table 1).

Reliability, Correlations, and Global Preparedness

Mean total scores were 64.77 (SD 16.40) for MAIRS-MS, 112.74 (SD 17.67) for AI attitudes, 45.90 (SD 8.71) for EPIS-TR, and 70.43 (SD 27.63) for AI anxiety. The α/ω coefficients for total scores were 0.957/0.957, 0.932/0.932, 0.930/0.929, and 0.958/0.959, respectively. Coefficients above 0.95 for MAIRS-MS and AI anxiety may reflect high consistency but also possible item redundancy. Subscale reliability estimates are provided in Supplementary Table S1 (Table 2).

After FDR correction, MAIRS-MS was associated with AI attitudes (ρ=0.32, 95% CI 0.11-0.50, q=0.018). AI attitudes were associated with EPIS-TR (ρ=0.26, 95% CI 0.02-0.47, q=0.038) and inversely with AI anxiety (ρ=-0.29, 95% CI -0.48 to -0.09, q=0.020). The CIs for the remaining correlations included zero. In the exploratory subscale analysis, three of 16 associations remained statistically significant after FDR correction: interest with MAIRS-MS cognition and ability, and belief in AI capabilities with MAIRS-MS ethics (Supplementary Material). This pattern indicates that the total-score association is not uniform across domains and may partly reflect overlapping favorable self-appraisals (Table 3).

Responses to the global preparedness item were: 1, n=5; 2, n=16; 3, n=35; 4, n=23; and 5, n=5. The item was associated with MAIRS-MS (ρ=0.38, 95% CI 0.16-0.57, p<0.001). Because both are self-ratings with similar wording, this finding indicates consistency between two self-reported indicators rather than independent validation or objective preparedness.

Exploratory Multivariable Associations and Diagnostics

The four-predictor model was statistically significant, F(4,79)=3.77, p=0.007, R²=0.160, adjusted R²=0.118. More positive AI attitudes were associated with higher MAIRS-MS scores (b=0.330, standardized β=0.356, 95% CI 0.112-0.548, p=0.004, q=0.014). Estimates for AI anxiety, EPIS-TR, and broadly defined previous training were imprecise and included zero; they should not be interpreted as evidence of no association (Table 4).

Diagnostic results were as follows: VIF 1.04-1.31; Shapiro-Wilk p=0.648; Breusch-Pagan p=0.109; maximum leverage=0.163; maximum absolute standardized residual=2.78; and maximum Cook’s distance=0.220. Six observations exceeded 4/N for Cook distance, six exceeded 2p/N for leverage, and two had |standardized residual|>2. Quadratic terms did not reach statistical significance at the 0.05 level (p=0.165 for attitudes, 0.062 for anxiety, and 0.054 for EPIS-TR), although the latter values warrant caution and do not establish strict linearity. The attitude estimate remained positive when the most influential observation was removed (b=0.417, 95% CI 0.199-0.634, p<0.001), with HC3 standard errors (b=0.330, 95% CI 0.064-0.596, p=0.016), after demographic adjustment (b=0.311, 95% CI 0.092-0.531, p=0.006), and after specialty adjustment (b=0.304, 95% CI 0.091-0.518, p=0.006). Full results of the sensitivity analyses are presented in Supplementary Material.

Discussion

In this exploratory single-center sample of residents from four specialties, AI attitudes were associated with self-reported medical AI readiness. The association was modest and remained positive across several sensitivity analyses. Anxiety, interprofessional identity, and broadly measured previous training had estimates that were too imprecise to establish either meaningful associations or their absence. The achieved sample had limited sensitivity to effects smaller than approximately |ρ|=0.30 and to heterogeneous subgroup differences.

The principal finding requires careful consideration of the constructs being measured. MAIRS-MS asks respondents to appraise their knowledge, ability, vision, and ethical preparedness, whereas the attitude scale includes willingness to learn, confidence in AI capabilities, and resistance. These constructs are not identical, but both can contain favorable self-appraisal, interest, and generalized positive responding. The subscale analysis suggested domain-specific rather than uniform overlap. Nevertheless, measurement from a common source, simultaneous administration, similar response formats, social desirability, acquiescence, and generalized positive-response tendencies may inflate associations among self-reported constructs.²⁸ The coefficient therefore does not show that changing attitudes will cause readiness to improve.

AI anxiety was inversely associated with attitudes but not independently with readiness. This is compatible with, but does not test, an indirect pathway through attitudes or acceptance.^{15, 16} Mediation and moderation were not examined because they were not prespecified and N=84 would yield unstable estimates. Similarly, EPIS-TR was related to attitudes but not independently to readiness. Interprofessional identity concerns collaborative belonging, commitment, and beliefs rather than technological competence; its role may be more relevant when AI changes shared workflows or accountability. Its interpretation is restricted because all participants were physicians and EPIS-TR was validated in multidisciplinary samples.

The training result is inconclusive. Only 15 participants reported any training, the item combined AI with health technology, no definition of qualifying training was provided, and content, duration, recency, depth, and assessment were not recorded. It therefore cannot evaluate curriculum effectiveness. Contemporary competency frameworks emphasize AI foundations, patient-centered use, ethics, privacy, security, bias, health equity, data science, application skills, and observable performance rather than exposure alone.^{29, 30} These domains should be evaluated in prospective curricula using knowledge tests, structured clinical tasks, critical-appraisal exercises, and behavioral outcomes.

Participants frequently reported using tools they perceived as AI-based, but the survey named general-purpose tools and did not identify applications or define “clinical decision support.” Information retrieval, drafting assistance, diagnostic recommendations, risk prediction, and regulated clinical decision-support software differ in validation, oversight, and risk. The present data therefore describe participants’ interpretations of their purposes of use; they do not establish clinical AI adoption, patient-care use, or safe performance.

The association between MAIRS-MS and the global preparedness item represents consistency between semantically related self-ratings, not validation against an external criterion. Physician self-assessment can diverge from observed competence.³¹ High perceived readiness could coexist with miscalibration, unrecognized knowledge gaps, automation bias, or unsafe use. Educational evaluation should therefore examine whether training improves knowledge, calibration, critical appraisal, ethical reasoning, and supervised performance, not merely whether self-reported readiness increases.

The four specialties were selected because of established local collaboration and shared involvement in mental health, child-focused, and longitudinal generalist care. This sampling frame enabled all eligible residents to be invited but limits transferability. Pediatrics has an emerging postgraduate AI education literature,¹² and residents across specialties identify humanistic, ethical, and responsibility-related concerns.²² However, the present sample did not include surgery, radiology, emergency medicine, laboratory specialties, or other technology-intensive fields. Differences in response rates, particularly the lower response rate in family medicine, could bias estimates if respondents differed in AI exposure or educational culture. Specialty adjustment did not materially alter the attitude coefficient, but specialty estimates were exploratory and imprecise.

Practical and Research Implications

The study did not test an educational intervention. The implications therefore concern priorities for curriculum development rather than demonstrated educational effects. Future programs should address core AI concepts and performance metrics; data quality and dataset shift; bias and health inequities; privacy and cybersecurity; hallucinations and automation bias; explainability and uncertainty; human oversight and accountability; distinctions between generative tools and regulated clinical systems; documentation of AI-assisted work; and patient communication. Competency frameworks should be translated into objective, observed assessments and longitudinal outcomes.^{29, 30}

Study Limitations

This study has several limitations. It was conducted at one center in four purposively selected specialties, with varying departmental response rates and no individual-level information about non-respondents. The cross-sectional, non-preregistered design precludes temporal or causal inference. All focal constructs were measured in the same online survey, creating common-method, social-desirability, acquiescence, and self-assessment biases. No knowledge test, observed clinical task, behavioral indicator, or social-desirability scale was used. The readiness and attitude measures were developed in educational populations, and EPIS-TR in multidisciplinary samples; high reliability does not establish construct validity in residents, and α/ω values above 0.95 may reflect redundancy. Training exposure was broad, heterogeneous, and defined by participants. AI-use purposes were not application-specific, “clinical decision support” was undefined, and patient-care use was not verified. The modest sample limited precision for small effects, specialty comparisons, and complex pathways; CIs that include zero should not be interpreted as proof of no relationship.

Conclusion

More positive AI attitudes were associated with higher self-reported medical AI readiness among residents from four specialties at one center. Partial construct overlap and common-method variance may contribute to this association. Associations with anxiety, interprofessional identity, and heterogeneous previous training were estimated imprecisely. Multicenter longitudinal and educational-intervention studies should distinguish AI categories and use objective performance measures to determine whether structured education improves safe, calibrated, and responsible AI practice.