Results of a survey among 155 primary care physicians in the mediX Switzerland network on attitudes, knowledge, ethical concerns, and managed care potential of AI in primary care.
Survey period: 2024 – approximately two years after the launch of ChatGPT (November 2022). The rapid development of AI since that time forms the context for the attitudes and knowledge of the surveyed physicians.
The 25% response rate falls within the expected range for online surveys among physicians. International comparative studies typically report response rates between 10% and 35% for similar surveys in primary care.
Comparison: Similar Studies
The central finding: while 69% of primary care physicians have a positive attitude toward AI, only 15% possess adequate knowledge. This 54 percentage point discrepancy (Cohen's h = 1.18) is the largest effect in the entire survey and indicates a considerable need for education.
Distribution on 5-point Likert scale. The gap between attitude (red) and knowledge (blue) is striking.
Effect size: Cohen's h = 1.18 (very large). The discrepancy persists across all subgroups.
Means with standard deviation (error bars). Mann-Whitney U, p < 0.001.
Cohen's d = 0.82 (large effect). LLM users show significantly higher AI knowledge.
Multiple responses permitted. Administrative applications dominate clearly.
The preference for administrative applications (80%) over diagnostic ones (49%) indicates a pragmatic approach.
Robustness of the knowledge-attitude discrepancy under different non-response scenarios
Even in the most conservative scenario, a substantial discrepancy persists.
The ethical concerns of physicians focus on two core themes: accountability for errors (66.7%) and lack of transparency (58.8%). The convergence with technical concerns (φ = 0.23) shows that trust in AI reliability is a prerequisite for ethical acceptance.
Multiple responses with 95% Wilson score confidence intervals
Comparison of two survey perspectives – notable convergence on accountability
Multiple responses with 95% Wilson score confidence intervals
The correlation (φ = 0.23, p < 0.01) between accountability concerns and transparency demands indicates that physicians view AI responsibility and technical traceability as inseparable.
50.7% of respondents see high to very high potential for AI in achieving managed care goals (M=3.47, SD=0.94). Treatment pathway support (58%) and quality measurement (49.3%) are identified as priority application areas.
Multiple responses with 50% reference line
Infrastructure concerns dominate
Acceptance among network members (30.0%) is rated as the lowest barrier – challenges are primarily infrastructural, not motivational.
Cohen's d = 0.40 (small to medium effect)
Donut chart of 5-point Likert distribution
The knowledge-attitude discrepancy (Cohen's h = 1.18) is the dominant finding. Swiss primary care physicians are fundamentally positive toward AI but lack sufficient knowledge for informed implementation. This carries the risk of both uncritical adoption and unfounded rejection.
The medical profession demands clear accountability frameworks before broader AI deployment. The convergence of ethical and technical concerns (φ = 0.23) shows that trust in AI reliability is a prerequisite for ethical acceptance. Regulatory clarity is urgently needed.
The strong correlation between general AI attitude and MC-specific potential (r = 0.54) suggests that investments in general AI education also promote acceptance of network-specific applications. Infrastructural hurdles – not lack of motivation – are the main barriers.
Develop practice-oriented curricula for primary care continuing education covering both technical fundamentals and ethical reflection. The high desire for training (81.9%) provides a favorable starting point.
The demand for legal frameworks (74%) and liability clarification (66.7%) requires proactive collaboration between the medical profession, regulators, and technology providers.
The largest implementation barrier (71.3%) is technical in nature. Investments in interoperable systems are a prerequisite for successful AI integration in physician networks.
Use administrative AI applications (80% approval) as an entry point. Accompanying research on effectiveness and acceptance in daily practice is essential.
In the interest of scientific transparency, the methodological limitations of this survey are openly presented below. Knowledge of these limitations is essential for appropriate interpretation of the results.
Convenience sampling with a response rate of 24.2–25.0% likely favors AI-interested physicians. However, sensitivity analysis shows that the knowledge-attitude discrepancy remains robust across all modeled scenarios (range: 2.5–54.2 percentage points).
Data originate exclusively from the mediX Switzerland network. Generalization to all Swiss primary care physicians or physicians outside managed care structures is not directly possible.
As an exploratory cross-sectional survey, the design does not allow causal conclusions. Observed associations could work in both directions. Longitudinal studies would be needed.
AI knowledge was measured by self-assessment, not objective knowledge tests. Dunning-Kruger literature suggests the actual knowledge gap may be even larger.
Despite anonymization, a bias toward socially desirable responses cannot be excluded, particularly for ethical questions.
88.4% of participants are from German-speaking Switzerland, 10.3% from Ticino, and only 1.3% from Romandie.
The questionnaire was not externally validated. Although based on existing literature, a formal psychometric validation is lacking.
The survey is explicitly positioned as hypothesis-generating, without a pre-registered study protocol.
This research project was conducted entirely independently and without financial support from public, commercial, or non-profit organizations.
The author declares that no conflicts of interest exist in connection with this research.
The datasets are not publicly accessible for data protection reasons. They can be reviewed upon justified request to the corresponding author.
According to the Swiss Human Research Act (HRA Art. 2), no ethics committee approval was required for this anonymized survey without collection of health data.
The exploratory cross-sectional surveys were conducted in 2024 within the mediX Switzerland network and follow STROBE and CHERRIES guidelines.
Exploratory cross-sectional survey via online questionnaire (SurveyMonkey). Email invitation to all registered physicians in the mediX Switzerland network (N≈620).
Descriptive statistics, Mann-Whitney U tests, Spearman correlations, chi-square tests, Wilson score confidence intervals. Effect sizes: Cohen's h, Cohen's d, Cramér's V, phi coefficient.
The complete methodology and all results are available in the preprint on medRxiv:
Read full text on medRxiv81.9% (127/155) – Spearman ρ = −0.02, p = 0.78
The following diagram illustrates the stepwise selection process from the target population to the analysable sample – analogous to a STROBE flow diagram.
5 questionnaires were excluded due to incomplete data (dropout before reaching core questions). All 150 remaining datasets were fully analysable.
54 of 150 participants (36.0%) reported regularly using LLM-based tools (e.g. ChatGPT). This subgroup serves as a comparison group in several analyses.
This study would not have been possible without the generous support of numerous individuals and institutions. The author wishes to express sincere gratitude:
Primary care physician and general internist in Zurich, regarded as one of the defining figures of physician-centered, coordinated care in Switzerland. For enabling the survey in the network and for organizational support that ensured smooth implementation.
Primary care physician and specialist in general internal medicine, combining clinical practice in primary care with guideline-based quality and health services research. For enabling the survey within the mediX Switzerland network and for valuable support in conducting the survey.
Internist and primary care physician specializing in ambulatory health services research, particularly in diabetes care, polypharmacy, and other primary care questions. For the initial exchange on the questionnaire and methodological suggestions that significantly contributed to the quality of the survey instrument.
Special thanks to all actively participating primary care physicians of the mediX network. Recognizing that these colleagues are extremely busy in their daily practice and have little free time, their willingness to participate in this survey is all the more appreciated.
Dr. med. Marco Vecellio is a specialist in general internal medicine and psychosomatic medicine. For over 15 years, he has been practicing in a Zurich group practice and closely observing the effects of digital transformation on primary care.
The present data were collected within the mediX Switzerland network. ORCID: 0009-0000-1772-5620.