Radiology Residents in Türkiye Trust AI as a Colleague, Not a Boss: Landmark National Survey Puts a Famous Attitude Scale to the Test

Artificial intelligence is moving into hospital reading rooms faster than most training programs can keep up with, and the people who will live with that transition most intimately are radiology residents — the doctors in the middle of a decade-long apprenticeship in image interpretation. A new multicenter study from Türkiye, published in BMC Medical Education, has now done something unusually rigorous with that tension: it took a widely used 14-item questionnaire measuring radiologists’ attitudes toward AI, translated and culturally adapted it into Turkish, and then stress-tested whether the instrument’s celebrated three-domain structure actually holds together when applied to a national cohort of 286 residents drawn from 25 institutions across 17 cities. The answer is a fascinating split verdict — one part of the scale survives intact, two parts crumble, and the residents themselves emerge as cautiously optimistic collaborators rather than anxious skeptics or uncritical enthusiasts.

The instrument at the center of the study is the ATRAI-14, short for Attitude Toward Radiology Artificial Intelligence, a 14-item questionnaire originally designed to capture how radiologists feel about AI implementation along three conceptual domains: Trust, Implementation Perspectives, and Hopes and Fears. Because residents’ attitudes are widely regarded as central to whether AI tools actually get adopted in clinical practice, the authors — Erdem Ozkan of Kastamonu Training and Research Hospital and Mehtap Balaban of Yıldırım Beyazıt University in Ankara — set out to adapt the scale into Turkish and examine whether its three-domain structure replicates in a different language and a different health system. The adaptation followed established cross-cultural methodology: the questionnaire was translated into Turkish and content-validated by a panel of eight expert radiologists, achieving an item-content validity index of 0.875, a figure indicating strong expert agreement that the translated items preserved the meaning of the originals.

The survey itself was national in scope and cross-sectional by design, administered as a web-based questionnaire to radiology residents across Türkiye. Psychometric evaluation was thorough and multi-pronged: confirmatory factor analysis to test whether the original three-factor structure reproduced itself, exploratory factor analysis to see what structure the data actually preferred, Cronbach’s alpha and McDonald’s omega to quantify internal consistency, item–total correlations to flag misbehaving questions, and criterion validity assessed against a visual analog scale of overall attitude. Group differences and predictors of the total score were examined with non-parametric tests and regression modeling. This is the full battery one would want before declaring a translated instrument fit for research use — and it is precisely this battery that produced the study’s most instructive negative results.

The headline psychometric finding is that the three-factor model fitted inadequately in the Turkish cohort. The comparative fit index came in at 0.82, well below conventional thresholds for acceptable model fit, and the root mean square error of approximation was 0.075, sitting at the edge of what researchers typically tolerate. In plain terms, the data from Turkish residents did not organize themselves into the three neat conceptual boxes the scale was designed around. Only the Trust domain reached borderline-acceptable internal consistency, with a Cronbach’s alpha of 0.69 — just under the 0.70 line most psychometricians treat as the minimum for group-level comparisons. The other two domains collapsed entirely: Implementation Perspectives returned an alpha of 0.07 and Hopes and Fears an alpha of 0.06, values so close to zero that they indicate the items within those domains share essentially no common signal. Two items even loaded negatively, pulling in the opposite direction from their intended constructs.

The authors are refreshingly candid about what this means, and their caution is worth emphasizing because it changes how the rest of the findings should be read. Because Implementation Perspectives and Hopes and Fears did not form coherent constructs in this population, findings touching those domains should be interpreted as observations about individual items rather than as measurements of the intended underlying dimensions, and they should not be used to compare groups. The total score, however, fared better: its criterion validity against the visual analog scale was acceptable, with a Spearman correlation of 0.56, meaning the overall number does track meaningfully with residents’ self-reported global attitude even if the sub-domains do not cohere. This is a genuinely useful contribution to the literature on cross-cultural instrument validation — a reminder that a scale validated in one linguistic and health-system context cannot simply be translated and assumed to measure the same things elsewhere.

So what do Turkish radiology residents actually think about AI? The mean total score was 20.99 with a standard deviation of 3.60, sitting 2.99 points above the scale midpoint of 18 — a modestly positive attitude overall. The item-level pattern is where the story gets vivid. On the collaborative side, 90.6% of residents said they would re-examine a study when AI disagreed with their reading, a striking endorsement of AI as a second pair of eyes worth taking seriously. Expectations about workload were similarly optimistic: 79.7% expected AI to reduce their workload. But the appetite for autonomy stops hard at the boundary of machine independence — only 22.0% of residents expressed comfort with AI operating autonomously. The residents want a colleague, not a replacement.

Not every sentiment was rosy. A majority — 60.8% — feared that AI would diminish the prestige of their profession, a concern that speaks to identity as much as to employment. And perhaps the most practically consequential number in the study is one of the smallest: only 21.7% of residents had actual access to AI tools at work. That gap between opinion and experience is the study’s quiet bombshell. Nearly four in five residents are forming attitudes about a technology they have almost never touched, which raises an obvious question about what those attitudes are actually measuring — informed judgment, or accumulated rumor, media coverage, and anxiety about the future of the specialty.

Equally telling is what did not predict attitudes. Residents’ scores did not differ by AI access, seniority, institution type, or research involvement. In a more mature adoption environment, one might expect hands-on users to diverge from outsiders, or senior residents with more clinical experience to view automation differently from juniors. The absence of any such differences in this cohort reinforces the impression that AI in Turkish radiology training remains largely an abstraction — everyone is weighing in from roughly the same distance. The authors are careful here too: they suggest that incorporating supervised, hands-on AI experience into residency training could help align expectations with practice, but they explicitly frame this as a hypothesis warranting direct evaluation rather than a conclusion drawable from cross-sectional data.

The study received no specific funding, was approved by the Kastamonu University Clinical Research Ethics Committee, and collected electronic informed consent from all participants. Its methodological honesty may prove to be its most exportable contribution. As AI attitude surveys proliferate across specialties and countries, this paper demonstrates both the necessity and the difficulty of proper psychometric validation: a translation that experts judged highly faithful (an item-content validity index of 0.875) still produced a scale whose two of three domains dissolved in a new population. For researchers hoping to track how the next generation of radiologists feels about AI, the message is clear — trust the Trust domain, treat the rest with caution, and above all, get residents their first real hands-on encounter with the algorithms everyone is already opining about.

Subject of Research: Cross-cultural adaptation and psychometric validation of the Turkish ATRAI-14 questionnaire measuring radiology residents' attitudes toward artificial intelligence

Article Title: Cross-cultural adaptation and psychometric evaluation of the Turkish ATRAI-14 among radiology residents: a multicenter national survey

Article References: Ozkan, E., & Balaban, M. (2026). Cross-cultural adaptation and psychometric evaluation of the Turkish ATRAI-14 among radiology residents: a multicenter national survey. BMC Medical Education, 26(1), Article 1653. https://doi.org/10.1186/s12909-026-10415-3

Image Credits: AI Generated

DOI: 10.1186/s12909-026-10415-3

Keywords: artificial intelligence, radiology, psychometrics, ATRAI-14, cross-cultural adaptation, medical education, radiology residents, survey research, factor analysis, Türkiye, attitude measurement, clinical training