Do Riyadh clinicians know AI better than they trust it?
A peer-reviewed survey of 400 physicians and dentists at three government university hospitals found 89.5% reported familiarity with clinical AI, while trust and intended adoption were only moderate. The result describes attitudes in selected Riyadh institutions—not whether any AI system improves care.
The short version
- What happened
- Among 400 clinicians, 89.5% reported familiarity with clinical AI; the sample comprised 254 physicians and 146 dentists.
- Who it affects
- Clinicians need evidence for a particular task and system; general AI familiarity is not a substitute for local validation.
- What remains unproven
- Cross-sectional associations cannot show that familiarity, profession, training or work setting caused trust or adoption intention.
Keep in mind
Cross-sectional associations cannot show that familiarity, profession, training or work setting caused trust or adoption intention.
Responses were self-reported; no usage logs, clinical decisions, diagnostic accuracy or patient outcomes were examined.
Purposive recruitment from three government university hospitals in Riyadh limits national and international generalisability.
Physicians and dentists may have evaluated different tools and clinical tasks.
Patients' views were not collected directly, and the instrument was not factor-analysed in this sample.
Source analysis · 1 linked source. Primary analysed source date: 13 October 2026.
Read the full report · 7 min The evidence, methods, implications and limitations
The answer: familiarity was high, but confidence was qualified
Clinicians in this Riyadh survey generally knew about clinical artificial intelligence, but awareness did not translate into unreserved trust. Of 400 respondents, 89.5% said they were familiar with clinical AI tools and 64% reported using a clinical decision-support system. Large language models were used daily by 10.5%, weekly by 22.8%, monthly by 25.2% and occasionally by 41.5%. Those figures show exposure, not dependence: the questionnaire did not audit systems, usage logs, decisions or patient records.
Attitude scores sat in the middle-to-positive range. The mean perception score was 64.8 out of 90, the trust score 19.8 out of 30, and intended adoption 18.5 out of 25. Those totals are more informative than a claim that clinicians simply accept or reject AI. A clinician can know a tool, recognise potential value and still withhold trust because the output is difficult to explain, locally unvalidated or unclear about responsibility. The paper therefore identifies an implementation question—not proof that reluctance is a problem to overcome.[1]
The study surveyed 254 physicians and 146 dentists
The authors used a cross-sectional online questionnaire in three government university hospitals: King Abdulaziz Medical City, King Saud University Medical City and King Abdullah Bin Abdulaziz University Hospital. They calculated that a population of 4,306 clinicians required about 354 responses at 95% confidence and a 5% margin of error, then increased the target to 400 for non-response and subgroup analysis. Invitations went through institutional email lists and professional WhatsApp groups to licensed clinicians who consented.
The final sample included 254 physicians and 146 dentists. Women made up 226 participants, 329 respondents were Saudi nationals and 365 worked full time. Professional levels ranged from 122 residents and 42 general practitioners to 79 specialists, 24 fellows and 133 consultants. The questionnaire adapted previously published instruments and reported Cronbach's alpha values from 0.880 to 0.952 for its subscales. High internal consistency shows that items within a scale moved together; it does not establish that the scale captures every relevant dimension of warranted clinical trust.[1]
Differences between groups may reflect context as much as profession
Physicians reported familiarity more often than dentists—92.9% versus 83.4%. Multivariable models also associated specialty, resident status and work setting with perception, trust or adoption scores. For example, physician specialty was associated with higher perception, trust and adoption scores after adjustment, while scores differed substantially between the three hospitals. The paper treats those patterns as signals for tailored implementation and training.
They should not be read as fixed characteristics of doctors, dentists or hospitals. Participants were not randomly assigned to institutions, professional roles or AI exposure. Hospitals may differ in digital systems, leadership, case mix, training and the exact tools clinicians have encountered. The study also measured associations at one point in time. It cannot establish that being a physician causes greater familiarity, that training raises trust, or that a particular level of trust produces safer adoption.[1]
What this means for clinicians and patients
For clinicians, the practical lesson is to separate generic AI literacy from confidence in a named system for a named task. Knowing how a language model works does not validate a diagnostic aid, and feeling comfortable with one administrative tool does not justify accepting another tool's clinical recommendation. Training should cover local performance, known failure modes, uncertainty, escalation and who remains accountable—not only how to operate the interface.
For patients, moderate trust can be protective when it prompts verification, but damaging if organisations mistake trust as a personal attitude problem and pressure staff to comply. Implementation teams should ask whether clinicians can inspect relevant evidence, challenge an output and document disagreement. Patient perspectives were not collected directly in this study. The paper therefore cannot tell us whether patients want these tools, understand their role, or experience different care when their clinician uses them.[1]
The evidence is local, purposive and self-reported
The sample came from three large, digitally advanced tertiary institutions in one city. The authors used purposive selection and voluntary online responses, which can favour people already interested in technology. Although the achieved sample exceeded the calculated minimum, a large denominator does not remove selection bias. The findings should not be projected onto community clinics, smaller hospitals, other Saudi regions or health systems with different infrastructure.
The questionnaire also relied on recall and self-description. Reported use does not show frequency in system logs, quality of use or whether a clinician followed an AI suggestion. Physicians and dentists may have been thinking about different products and tasks when answering the same questions. The authors did not perform exploratory or confirmatory factor analysis in this sample, and the study did not deeply operationalise explainability, regulation or ethics. These limits narrow the result to perceived readiness rather than real-world safety or benefit.[1]
What would change the assessment
Confidence would rise with representative sampling across Saudi regions and care settings, followed by longitudinal work that links named training or governance changes to later behaviour. The most useful evaluations would identify the tool and task, compare intended with observed use, record overrides and escalation, and measure diagnostic quality, workload and patient outcomes. Patient interviews should be analysed separately rather than inferred from clinicians.
A stronger implementation case would also require prospective trials showing that a specific system improves care without shifting hidden work onto clinicians or widening inequity. If attitudes remain moderate after transparent local validation, usable explanations and clear accountability, that would suggest concerns are not solved by literacy alone. If trust becomes appropriately calibrated—rising for reliable tasks and falling for weak ones—that is more valuable than maximising a single trust score.[1]