Introduction
Background
The use of artificial intelligence (AI) is rapidly expanding across the medical field, including diagnosis, prognostic prediction, treatment support, and administrative automation, and is emerging as a core technology in modern healthcare systems [1,2]. However, medical AI involves more than the introduction of new technology; it raises complex policy and ethical challenges related to reliability, accountability, safety, and institutional coherence [3–5]. Because medical AI can directly affect patient health and survival, it requires a higher level of validation and broader social consensus than AI applications in other industries [6,7]. In this context, perceptions of medical AI play a decisive role in shaping technological adoption and policy design. Medical professionals tend to evaluate medical AI primarily through the lens of patient safety and legal and ethical responsibility, whereas non-medical professionals are more likely to emphasize structural considerations, such as technological scalability, industrial competitiveness, and data infrastructure. These differing perceptions represent an important factor in determining whether medical AI policies prioritize safety-oriented regulation or emphasize innovation and diffusion.
Nevertheless, existing studies have largely focused on evaluating the technical performance of medical AI systems and analyzing individual applications [8,9]. Comparative studies that systematically examine differences in perceptions between medical and non-medical professionals remain limited. In particular, few qualitative investigations have explored how perceptions regarding reliability, responsibility, educational needs, research and development (R&D) directions, and legal, institutional, and policy priorities are structured differently across expert groups. Consequently, misalignment between clinical demands and industrial or institutional perspectives may recur during policy-design processes.
Objectives
To address this gap, this study aimed to qualitatively compare and analyze perceptions of medical AI among medical and non-medical professionals across multiple domains, including perceived performance, technical limitations, reliability and responsibility, educational needs, R&D directions, and legal, institutional, and policy priorities. The study sought to systematically elucidate the underlying perception structures and key issues characterizing each group and to provide a conceptual foundation for future medical AI policies that more effectively balance safety and innovation.
Methods
Ethics statement
This study complied with the ethical standards of the Declaration of Helsinki and was approved by the Institutional Review Board of The Catholic University of Korea, Seoul St. Mary’s Hospital (KC25QISI0819). All participants received a detailed explanation of the study objectives, procedures, and data use. Written informed consent was obtained based on voluntary participation and assured anonymity. Ethical principles were strictly observed, including the protection of participants’ anonymity and freedom of expression.
Research team and reflexivity
(1) Interviewer/facilitator: Hun-Sung Kim. (2) Credentials: MD, PhD. (3) Occupation: Professor at the College of Medicine, The Catholic University of Korea. (4) Gender: male. (5) Experience and training: The interviewer has extensive experience in qualitative research, including focus group interviews and consensual qualitative research. (6) Relationship with participants: Before the interviews, the researcher established rapport with participants by providing a detailed explanation of the study process and background. Participants were fully informed of the interviewer’s institutional affiliation, professional background, and the specific objectives and methods of the study. As a professional with substantial expertise and academic interest in medical AI, the interviewer shared a similar academic and clinical background with many participants, which facilitated in-depth and open discussion.
Study design
Theoretical framework
This pilot qualitative comparative study examined perceptions of medical AI among medical and non-medical professionals. Because medical AI represents a sociotechnical phenomenon in which clinical application [10], technological development, and policy design are closely interconnected, structural differences in perception cannot be adequately captured using single survey items or quantitative indicators. Accordingly, a qualitative design was adopted to explore how expert groups with distinct roles and responsibility structures interpret and construct meaning around medical AI. In particular, the study conceptualized differences in perception not merely as attitudinal variation, but as differences in interpretive frames shaped by professional roles and responsibilities. To achieve this objective, a qualitative analytic strategy was employed that enabled systematic comparison between medical and non-medical expert groups.
Participant selection
Participants were experts with interest and experience in medical AI and were recruited through an open call based on voluntary participation. Information regarding the study purpose and procedures was disseminated through professional networks related to medical AI. The first 30 respondents who expressed clear willingness to participate were selected on a first-come, first-served basis. To enable comparison of different roles and responsibility structures, the sample comprised 15 medical and 15 non-medical professionals, ensuring equal group sizes to prevent disproportionate influence from a single group. As this was a pilot qualitative study, the maximum sample size was predetermined at 30 participants, balancing feasibility within the study timeline with the depth required for qualitative analysis. Participant selection did not aim for random sampling or statistical representativeness. Consistent with qualitative research objectives, recruitment focused on experts willing to engage in reflective discussions on medical AI. All participants had professional experience related to medical AI in at least one domain, including clinical practice, research, technological development, policy, industry, or institutional design.
Setting
Focus group interviews were conducted between October and November 2025. To accommodate participants’ geographic locations and professional schedules, sessions were held either in person in a private conference room at an academic medical center or via real-time videoconferencing using a secure digital platform (Zoom Video Communications). All environments were carefully managed to ensure confidentiality and minimize external distractions, thereby providing a secure setting for participants to share their professional experiences, perceptions, and concerns regarding medical AI.
Data collection
Data were collected over approximately 2 months. As this was a pilot study, the procedure was specifically designed to explore and conceptualize the underlying interpretive frames shaping perceptions of medical AI. Data were gathered through mixed focus group interviews that included both medical and non-medical professionals to identify perceptions emerging from intergroup interaction. Interviews followed a predefined semi-structured guide, and each session lasted approximately 60–90 minutes. To maintain rigor and minimize external influence, only the interviewer and participants were present during both in-person and online sessions. The interviewer moderated discussions to prevent dominance by any single group and to facilitate balanced participation. With participants’ prior consent, all sessions were audio-recorded (and video-recorded for online sessions) and transcribed verbatim for subsequent inductive thematic analysis.
Qualitative data analysis
Data analysis was conducted using inductive thematic analysis supplemented by procedures enabling systematic cross-group comparison. Interview transcripts were repeatedly reviewed to identify meaningful units reflecting perceptions, evaluative criteria, concerns, and expectations related to medical AI. The professional background associated with each statement was documented to facilitate comparison between groups. Statements with similar meanings were grouped into analytic categories, which were subsequently organized into higher-level themes. Analytical emphasis was placed on the criteria and reasoning through which medical and non-medical professionals interpreted shared issues—particularly those related to performance, risk, autonomy, and responsibility—rather than on the frequency of specific topics. Attention was also given to interactional processes through which contrasting views were articulated or reconciled during discussion. Multiple researchers participated in the analytic process, and final themes were refined through iterative discussion and consensus.
Results
The qualitative analysis indicated that medical and non-medical professionals shared a generally positive assessment of the overall value and potential of medical AI. However, clear differences emerged in the criteria used to interpret performance, recognize limitations, and formulate expectations regarding systems and policies. These differences did not appear to reflect individual familiarity with technology; instead, they seemed to reflect distinct interpretive frames shaped by differing roles and responsibility structures in relation to medical AI (Figure 1).
Perceptions of medical AI performance
Medical and non-medical professionals agreed that medical AI has developed rapidly in recent years and has generated meaningful outcomes in clinical practice and related industries. However, differences emerged in the criteria used to evaluate medical AI performance (Table 1). Medical professionals primarily assessed medical AI in terms of clinical usefulness and workflow efficiency, emphasizing tangible benefits such as reduced burden of repetitive tasks and enhanced diagnostic support in areas including medical image interpretation, electrocardiogram analysis, and clinical documentation. In contrast, non-medical professionals evaluated medical AI performance in terms of technological maturity and scalability. Although they acknowledged that image-based AI systems have approached—or, in some cases, exceeded—expert-level performance, they emphasized persistent limitations in data quality and generalizability, particularly for non-image and multimodal AI applications. This perspective reflects an understanding of medical AI as a mid- to long-term technology platform rather than a solution defined primarily by short-term clinical performance.
Awareness of key limitations of medical AI
Both groups repeatedly identified limitations of medical AI, although their focal concerns differed. Medical professionals described reliability and accountability as the most critical barriers, highlighting limited explainability, risks of hallucinations, and incomplete integration with electronic medical record systems. While acknowledging the utility of AI as an assistive tool, medical professionals consistently emphasized that medical AI cannot replace human clinical assessment or ultimate professional responsibility. In contrast, non-medical professionals focused on structural and systemic challenges, including data quality issues, lack of standardization, and constraints on generalizability. Inconsistencies in data formats, incomplete labeling, and data bias were frequently cited as factors that undermine stable AI performance and limit industrial scalability and global competitiveness. High infrastructure costs and regulatory barriers were also identified as practical constraints on the diffusion of medical AI technologies.
Perceptions of medical AI education
Medical and non-medical professionals strongly agreed on the need for education related to medical AI, although differences emerged in educational focus (Table 2). Medical professionals emphasized that education should extend beyond operational training to include an understanding of the principles and limitations of medical AI, as well as patient safety and ethical and legal responsibility. They stressed the importance of preparing clinicians to identify potential errors in AI systems and to understand the implications for clinical accountability. In contrast, non-medical professionals approached medical AI education from the perspective of enhancing utilization capabilities and interdisciplinary collaboration. They expressed concern that insufficient AI literacy among clinicians could limit potential benefits or negatively affect healthcare quality, and they emphasized the need for systematic and continuous educational curricula. Across both groups, there was shared recognition that education on emerging technologies, including large language models, should be expanded across medical schools and healthcare institutions.
Perceptions of R&D and policy directions
Regarding R&D and policy, medical and non-medical professionals expressed both converging and diverging perspectives (Table 3). Medical professionals emphasized that R&D should prioritize addressing concrete clinical needs and identified safety enhancement, explainable AI, clinical workflow integration, and rigorous validation as key priorities. From a policy perspective, they regarded flexible regulatory frameworks and clearly defined responsibility-sharing structures as essential for broader adoption. In contrast, non-medical professionals framed R&D priorities in terms of industrial competitiveness and global scalability, highlighting the importance of data infrastructure, standardization, regulatory rationalization, and appropriate reimbursement mechanisms for long-term sustainability. Despite these differences, both groups agreed that medical AI policy should adopt an integrated approach that aligns trust, education, research, responsibility, and governance rather than positioning patient safety and technological innovation as opposing goals.
Perception of the medical AI classification framework
Medical and non-medical professionals agreed on the need for systematic classification of medical AI, recognizing it as a complex technology encompassing multiple dimensions, such as clinical purpose, data type, technological structure, and level of autonomy. Both groups noted that single-axis classification schemes are insufficient for effective policy and institutional design. However, clear differences emerged in how each group conceptualized the functions of classification systems. Medical professionals primarily viewed classification as a tool for aligning AI applications with clinical workflows and patient journeys, favoring stage-based approaches reflecting prevention, diagnosis, treatment, and follow-up. This perspective emphasized defining the scope of AI intervention and associated clinical responsibilities while also acknowledging overlapping boundaries between clinical stages that necessitate multidimensional classification. In contrast, non-medical professionals regarded classification systems as a shared framework for policy, industry, and regulation. They emphasized implementation-oriented dimensions, such as data type, technological architecture, and autonomy level, to ensure consistency across development, approval, reimbursement, and market entry. In particular, emerging technologies such as multimodal and generative AI were described as requiring dynamic, continuously updatable classification systems rather than fixed taxonomies.
Perceptions of autonomy and responsibility structures in medical AI
Both groups identified the level of autonomy in medical AI and associated responsibility structures as critical issues, although differences in emphasis were evident (Table 4). Medical professionals expressed concerns that increasing AI autonomy is associated with greater clinical risk and increased uncertainty regarding liability. They generally regarded partially autonomous, human-in-the-loop configurations—where humans retain final decision-making authority—as the most realistic and appropriate model in the current healthcare environment. Regarding fully autonomous medical AI, medical professionals emphasized that applications should remain limited unless robust institutional mechanisms are established to ensure patient safety and legal accountability, regardless of technical feasibility. Non-medical professionals also acknowledged the importance of responsibility and accountability but articulated a more gradual and strategic view of autonomy. Rather than uniformly restricting autonomy levels, they emphasized differentiated approaches based on clinical purposes and risk. In particular, expanding autonomy in low-risk contexts or repetitive decision-making tasks was viewed as a way to improve efficiency and accessibility. Some participants suggested targeted institutional pilot programs to explore such applications under controlled conditions.
Perceptions of governance and continuous management of medical AI
Both medical and non-medical professionals emphasized the need for continuous management, updating, and verification of medical AI systems (Table 5). Because AI performance can evolve over time, participants broadly agreed that regulatory models focused solely on premarket approval—similar to those for conventional medical devices—are insufficient. Instead, governance frameworks incorporating ongoing validation and post-deployment monitoring were described as essential. Medical professionals emphasized the importance of hospital-level internal validation mechanisms and clear delineation of responsible parties, arguing that healthcare institutions should play an active role in continuously evaluating real-world clinical performance rather than serving solely as end users. In contrast, non-medical professionals highlighted the need for nationally standardized validation frameworks and public monitoring platforms, viewing trust in medical AI as dependent on multilayered governance involving developers, healthcare institutions, and governmental authorities.
Discussion
Key results
This qualitative analysis suggests that both medical and non-medical professionals hold generally positive views about medical AI, but they evaluate it using different criteria. Medical professionals emphasized practical usefulness and fit within day-to-day clinical workflows, whereas non-medical professionals focused on technological maturity and scalability. For example, medical professionals often raised concerns about how AI performs in clinical decision-making and whether it disrupts established routines, whereas non-medical professionals emphasized whether the technology is sufficiently developed and can be implemented at scale.
These differing emphases were also reflected in perceived educational needs, R&D priorities, preferred classification approaches, acceptable levels of autonomy, and perspectives on regulation and oversight. Overall, the findings suggest that each group’s interpretive frame is shaped by professional roles and responsibility structures.
Interpretation/comparison with previous studies
Divergent cognitive frames between clinical and system/policy perspectives
The findings indicate that, although both groups shared a strong consensus regarding the overall value and potential of medical AI, clear differences emerged in the criteria used to interpret technological performance, recognize limitations, and structure expectations for systems and policies. Importantly, these differences do not appear to be attributable to individual levels of technological familiarity or access to information; instead, they reflect distinct interpretive frames shaped by differing roles and responsibility structures associated with medical AI.
Medical professionals predominantly perceived medical AI as a high-risk clinical tool intended to support, rather than replace, clinical assessment. Their evaluations emphasized patient safety, reliability, explainability, and legal accountability as central criteria. This perspective may be interpreted as a pragmatic stance grounded in clinical experience, in which malfunctions or errors in medical AI can have immediate and substantial consequences for individual patients. These findings are consistent with previous studies reporting that, although medical experts acknowledge potential benefits of medical AI, they repeatedly identify limited explainability and ambiguous responsibility as major barriers to acceptance [11–13]. In particular, studies on AI-based medical imaging and clinical decision-support systems have consistently shown that clinicians tend to position AI as an assistive instrument rather than an autonomous decision-maker [14,15]. In contrast, non-medical professionals tended to conceptualize medical AI as a technology platform capable of reorganizing the broader healthcare system, emphasizing structural factors such as technological scalability, system integration, data infrastructure, and industrial sustainability. This perspective aligns with technology- and policy-oriented literature framing medical AI as a driver of system-level innovation aimed at improving efficiency and cost structures in healthcare. Studies in the policy and governance literature have emphasized data standardization, institutional predictability, and regulatory coherence as critical conditions for the diffusion of medical AI, reflecting a cognitive orientation similar to that expressed by the non-medical participants in this study.
As illustrated in the conceptual framework presented in this study, these divergent perspectives can be structured as a contrast between a “clinical frame,” which interprets medical AI primarily through the lens of clinical risk and professional responsibility, and a “system/policy frame,” which emphasizes system design, regulatory architecture, and industrial organization. Although prior research has largely examined attitudes toward medical AI or levels of acceptance within individual professional groups [16], this study contributes by qualitatively comparing how groups with different roles and responsibility structures interpret the same technology through fundamentally different frames.
These findings suggest that discussions of medical AI focused narrowly on technical performance or isolated regulatory issues risk a disconnect between clinical adoption and policy implementation. Policies reflecting only safety-centered demands may slow innovation, whereas those driven solely by industrial and technological imperatives may fail to secure trust and acceptance in clinical settings. The results empirically support the tension frequently highlighted in the medical AI governance literature.
Classification system differences between clinical purposes and technical implementation approaches
This study identified clear differences between medical and non-medical professionals in perceptions of medical AI classification systems. Medical professionals favored classification schemes centered on clinical purpose and the patient journey as most reflective of real-world practice. In contrast, non-medical professionals emphasized multidimensional classification systems incorporating technical implementation structures, data types, and levels of autonomy as essential tools for policy and industrial design. These findings underscore the inherent complexity of medical AI and support prior discussions emphasizing that its multidimensional nature cannot be adequately captured using a single classification standard [17]. International regulatory trends are consistent with these observations. For example, the European Union’s AI regulatory framework designates medical AI as a high-risk domain while applying differentiated requirements based on intended use and risk level rather than treating medical AI as a uniform category [18]. Similarly, the U.S. Food and Drug Administration has proposed regulatory approaches that account for the learning and adaptive characteristics of medical AI systems rather than subsuming them entirely under traditional medical device classifications [19]. These approaches are structurally consistent with the multidimensional classification perspective emphasized by the non-medical professionals in this study.
Autonomy and responsibility structures contrasting phased introduction with strategic differentiation
Differences in perceptions of autonomy levels and associated responsibility structures have important implications for future regulatory design. Medical professionals generally preferred phased introduction of medical AI based on human-in-the-loop models, emphasizing that increasing autonomy is closely associated with increased clinical risk and legal uncertainty. In contrast, non-medical professionals articulated a more strategic and differentiated view, suggesting that autonomy could be selectively expanded depending on the clinical purpose and risk profile of specific applications. This perspective is consistent with critiques in the literature that dichotomous debates framing medical AI as either “autonomous” or “non-autonomous” are insufficient for effective regulation [20,21]. Accordingly, governance studies have emphasized conceptualizing autonomy in medical AI as a continuous spectrum and adopting risk-based regulatory models that adjust oversight and responsibility according to context [22,23]. Within this framework, the findings of this study suggest that medical AI regulations should be designed in a differentiated manner, accounting for clinical purpose, risk level, and technological characteristics rather than relying on uniform autonomy thresholds or single-axis classification systems. Therefore, clinical safety and accountability concerns emphasized by medical professionals and the system-level design and policy feasibility considerations emphasized by non-medical professionals should be understood as complementary components of medical AI governance rather than competing demands. By empirically demonstrating the structure of these perspectives, this study provides a foundation for exploring integrative policy designs that reconcile safety and innovation in medical AI.
Educational approaches comparing ethics and safety focus with utilization and collaboration priorities
Analysis of perceptions regarding medical AI education showed strong consensus regarding the need for education while also highlighting clear differences in educational focus. Medical professionals emphasized ethics, legal responsibility, patient safety, and understanding limitations of medical AI as central educational components. In contrast, non-medical professionals prioritized utilization capabilities, interdisciplinary collaboration, and familiarity with emerging technologies such as large language models. These findings suggest that medical AI education should not be approached through a single uniform curriculum, but should instead involve differentiated educational designs tailored to distinct roles and responsibility structures. These differences align with existing literature on medical AI education [24,25]. Previous studies have emphasized that AI education for medical professionals should extend beyond operational training to include ethical reasoning, responsibility awareness, and understanding of clinical limitations [26,27]. In parallel, policy- and technology-oriented work has argued that effective diffusion of medical AI depends on clinicians’ ability to understand underlying principles and collaborate effectively with developers and system designers [28]. The present study adds empirical support for these discussions by demonstrating that such priorities are reflected in stakeholder perceptions.
Governance and continuous management showing shared recognition of dynamic oversight needs
Notably, both medical and non-medical professionals agreed on the importance of continuous management and post-deployment monitoring of medical AI. This shared recognition reflects an understanding of medical AI not as a fixed medical device, but as a dynamic technology whose performance and characteristics can evolve through ongoing updates. Such perceptions align with international governance literature questioning the adequacy of static premarket approval models derived from traditional medical device regulation for medical AI [29]. Accordingly, studies have argued that medical AI requires governance frameworks extending beyond initial approval to include continuous performance validation, error monitoring, and clearly defined responsibility structures throughout the deployment lifecycle [30]. Taken together, these findings indicate that trust in medical AI cannot be secured through technical performance alone, but must be supported through an integrated approach encompassing education, accountability, and governance. This study further suggests that demands for education and continuous management are not unilateral claims of a single stakeholder group, but reflect shared priorities across groups with differing roles and responsibilities.
Therefore, the findings support the need for dynamic, multilayered governance models capable of accommodating both clinical safety concerns and system-level implementation requirements. Through qualitative comparative analysis, this study provides insights into differences in perceptions of medical AI between medical and non-medical professionals.
Limitations
First, because of the qualitative study design, the findings are not statistically generalizable [31]. Participants were experts engaged in medical AI to varying degrees, and their perceptions may not represent those of the broader population of clinicians or non-expert groups. However, the primary aim of this study was not to estimate the distribution of attitudes, but to explore underlying perception structures and interpretive frames that shape policy and institutional design; therefore, a qualitative approach was appropriate. Second, the relatively small sample size was a limitation. Although a sample of 15 participants in each group was sufficient to support in-depth qualitative analysis, it limited the ability to examine heterogeneity within groups—for example, differences across clinical specialties among medical professionals or across sectors (e.g., industry, policy, and research) among non-medical professionals. Accordingly, medical and non-medical participants were analyzed as aggregated groups, and more nuanced intra-group variation warrants investigation in future studies. Finally, although this study focused on professional perceptions of medical AI, other stakeholders—including patients, the general public, hospital administrators, and insurers—also play critical roles in policy formation and technology acceptance. The perspectives of these groups were beyond the scope of this study, which limits the interpretation of the findings. Future research could benefit from mixed-methods designs incorporating a broader range of stakeholders to examine the multilayered governance structure of medical AI and to further test the interpretive frames identified in this study.
Implications
Nevertheless, this study makes academic and policy contributions by conceptualizing differences in perceptions of medical AI not as simple pro–con positions or attitudinal variation, but as differences in interpretive frames shaped by role and responsibility structures. Through qualitative comparison, the study demonstrates that medical and non-medical professionals interpret the same technology through distinct problem definitions and policy expectations. In doing so, it complements contextual dimensions of medical AI acceptance that are not well captured in many quantitative perception studies. In particular, by empirically linking stakeholder perceptions to discussions of medical AI classification systems, autonomy, responsibility structures, education, and governance, this study highlights that classification systems are not merely technical arrangements, but also function as instruments for policy design, regulation, and education. From a policy perspective, the findings suggest that diffusion of medical AI requires an integrated strategy connecting trust, education, research, responsibility, and governance rather than positioning patient safety and innovation as opposing goals. The study further emphasizes that medical AI policy should be understood as a process of coordinating and integrating diverse stakeholder frames rather than privileging the perspectives of any single group. Finally, this study aligns with international scholarly discussions that frame medical AI as a sociotechnical system and provides a conceptual foundation for future quantitative and mixed-methods research.
Conclusion
This pilot qualitative study explored perceptions of medical AI among medical and non-medical professionals using qualitative comparative analysis. The findings indicate that, although both groups agree on the potential value of medical AI, they interpret its meaning, risks, and policy priorities through distinct interpretive frames. Medical professionals tend to evaluate medical AI primarily in terms of clinical safety and professional responsibility, whereas non-medical professionals emphasize technological scalability and institutional and industrial sustainability. These differences help explain recurring tensions in medical AI policy debates and institutional design. The findings suggest that medical AI policies should not be developed by dichotomously opposing safety and innovation, but by coordinating and integrating diverse stakeholder frames. By structurally explicating differences in perceptions of medical AI, this study provides conceptual and empirical foundations for future quantitative, mixed-methods, and cross-national comparative studies on medical AI governance.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
