This study was conducted in China and consists of 2 parts. The first part used a grounded theory approach to analyze and construct an ethical risk framework for the clinical application of medical AI [24]. The second part involved the preliminary design of a questionnaire based on this framework, which is then revised and refined through focus group discussions [25], forming the final survey questionnaire to conduct the subsequent survey [26].

A multiple linear regression model was used to identify key predictors of risk perception. Independent variables were selected based on literature review and clinical relevance, including years of service, hospital level, familiarity with AI, training experience, and the establishment of medical institution AI ethics review procedures. For categorical predictors, the following reference categories were used: years of service (>15 y), hospital level (secondary and primary hospitals), familiarity with AI (familiar), training experience (none), and establishment of AI ethics review procedures (no). Significant variables were incorporated into the model, and results were detailed with β coefficients, 95% CIs, and P values. All analyses were performed using IBM SPSS Statistics 26.

All participants provided electronic informed consent before the study began. This study involves an anonymous online questionnaire approved by the Ethics Committee of China-Japan Friendship Hospital (approval 2024-KY-254). Instead of providing physical signatures, participants reviewed an informed consent statement on the survey’s opening page. This statement detailed the study’s purpose, its voluntary nature, confidentiality protocols, and the right to withdraw. Respondents had to click “I have read and agree to participate” to access the questionnaire. No compensation was offered to participants, as the survey was voluntary, posed minimal risk, and placed no financial burden on respondents. Data collected were anonymized, encrypted, and strictly limited to academic research purposes.

Based on the grounded theory approach, this study ultimately distilled a theoretical framework with “types of ethical risks in medical AI” as the core category. This framework covers 5 main categories, namely physiological risks, psychological risks, data and privacy risks, social risks, and economic and sustainability risks, along with their 15 subcategories (Table 1). Specifically, physiological risks refer to potential or actual physical harm to patients resulting from misdiagnosis or mistreatment due to algorithmic biases, limited data samples, or technical malfunctions in the process of AI-assisted diagnosis and treatment [28]. Psychological risks involve patients’ distrust and anxiety toward AI-driven decisions, as well as technical anxiety among health care providers caused by overreliance on AI [29,30]. Data and privacy risks denote security issues such as the leakage, misuse, or unauthorized secondary use of patient private data during data collection, storage, transmission, and usage by medical AI systems [31,32]. Social risks encompass multidimensional socioethical issues arising from the application of medical AI, including equity concerns stemming from uneven allocation of AI resources across regions [33]; the physician-patient trust crisis caused by the opacity of AI decisions, and the weakened decision-making autonomy of both physicians and patients due to the intervention of AI technology [34]; ambiguity in liability attribution among developers, medical institutions, and users when medical errors occur [35]; and occupational impacts caused by AI-driven substitution pressure on physicians’ roles [36]. Economic and sustainability risks mainly refer to the financial burden placed on medical institutions due to high costs related to the development, procurement, and maintenance of medical AI, as well as sustainability challenges, such as excessive energy consumption, resource waste, and environmental pollution resulting from rapid technological iteration [37].

This study included 600 physicians, and 0 people refused to participate. The characteristics of the sample group are as presented in Table 2. In terms of sex, males accounted for 46% (n=276) and females accounted for 54% (n=324); regarding age groups, physicians aged 40 years and younger accounted for 51% (n=306), those aged 41‐50 years accounted for 34.5% (n=207), while physicians aged 51 years and older accounted for only 14.5% (n=87). The majority worked in tertiary hospitals (n=375, 62.5%), while 37.5% (n=225) were affiliated with secondary and primary hospitals. Physicians with a bachelor’s degree or below accounted for 61% (n=366), and those with a postgraduate degree accounted for 39% (n=234). Regarding professional titles, resident physicians account for 14.5% (n=87), attending physicians account for 39.8% (n=239), associate chief physicians account for 30.8% (n=185), and chief physicians account for 14.8% (n=89). In terms of years of practice, physicians with 15 years or less of practice accounted for 54.3% (n=326), and those with more than 15 years of practice accounted for 45.7% (n=274). Most physicians reported a moderate to low level of familiarity with medical AI, 21.7% (n=130) indicated they were familiar, 47.7% (n=286) were neutral, and 30.7% (n=184) were unfamiliar. Additionally, 21% (n=126) had received specialized AI training, 33.8% (n=203) had self-studied topics related to medical AI, and 45.2% (n=271) had received no training at all. In clinical practice, 55.2% (n=331) had used AI, while 44.8% (n=269) had not. Furthermore, the development of ethical governance systems appears significantly underdeveloped; only 16.8% (n=101) reported that their hospital had established standard ethical review procedures for medical AI, 37.5% (n=225) stated that no such procedures were in place, and 45.7% (n=274) were unaware of whether such procedures existed.

This study used a structured questionnaire survey to systematically analyze physicians’ perceptions of risks in medical AI applications across 5 dimensions—physiological risks, psychological risks, data and privacy risks, social risks, and economic and sustainability risks (Figure 1).

Regarding the perception of physiological risks (C1), for Q1 “AI diagnostic errors delay treatment,” 51.3% (n=308) of physicians expressed agreement, while only 20.5% (n=123) disagreed. For Q2 “AI treatment errors endanger lives,” 52.2% (n=313) of physicians indicated apprehension, 21.3% (n=128) did not believe that AI would pose such risks, and nearly one-third remained neutral.

Perceptions of psychological risks (C2) revealed a notable asymmetry between “patient psychological anxiety” and “physicians’ technical anxiety.” Regarding Q3 (“AI replacing clinical roles causes anxiety”), 28.5% (n=171) of physicians expressed agreement, while 41.8% (n=251) disagreed and 29.7% (n=178) remained neutral. As for Q4 (“doctors’ confusion over AI logic creates stress”), 44.2% (n=265) of physicians acknowledged such risk, whereas 55.5% (n=335) denied it. In contrast, issues related to patient psychological anxiety received broader consensus. Specifically, 56.7% (n=338) of physicians agreed with Q5 (“Patients’ confusion over AI logic causes concern”), and an equal proportion of 56.7% (n=340) endorsed Q6 (“Patients worry about AI result accuracy”).

Data and privacy risks (C3) emerged as the dimension with the highest consensus among physicians. A majority expressed agreement with Q7 (“AI data collection violates privacy”; n=404, 67.3%), Q8 (“AI data storage risks leakage”; n=437, 72.8%), and Q9 (“Patient data used commercially infringes rights”; n=456, 76%). Neutral attitudes were reported by 18.2% (n=109), 16% (n=96), and 14.7% (n=88) of respondents, respectively, while disagreement was less than 14.5% (87/600) across all items. These results indicate a high level of shared concern regarding such risks among the physician community.

Regarding social risks (C4), Q15 (“AI misdiagnosis leads to ambiguous accountability”) received the highest agreement rate at 75.7% (n=454), with only 9.7% (n=58) expressing disagreement. Physician-patient trust crisis also drew significant attention, with agreement rates of 68.3% (n=419) for Q13 (“AI reduces physicians’ patient emotional connection”) and 72.3% (n=434) for Q14 (“AI uncertainty raises doubts among patients/families”). Notable divergence was observed concerning Q12 (“AI threatens physicians’ job security”), with 41% (n=246) in agreement and 30.7% (n=184) in disagreement, while 28.3% (n=170) remained neutral. Perceptions of fairness risks revealed a mix of concern and skepticism. The agreement rates for Q10 (“AI algorithm flaws cause healthcare bias”) and Q11 (“uneven AI adoption widens digital divide”) were 54.7% (n=328) and 63.3% (n=380), respectively. Most physicians agreed with autonomy erosion (Q16 and Q17), with agreement rates of 49.5% (n=297) and 55.8% (n=335), respectively.

Perceptions of economy and sustainability risks (C5) were the lowest among all dimensions. Only Q19 “AI misuse leads to medical resource waste” received agreement from over half of the respondents (n=336, 56%). The other issues have not yet garnered widespread attention among physicians; Q18 “AI adoption raises patient costs,” Q20 “AI medical waste harms the environment,” and Q21 “AI equipment increases energy demands” received agreement rates of merely 44% (n=264), 38.5% (n=231), and 49.2% (n=295), respectively.

This study used a multiple regression model to analyze the effects of years of practice, familiarity with AI, training experience, hospital level, and the establishment of standard AI ethics review procedures on physicians’ risk perceptions across the 5 major dimensions—physiological, psychological, data and privacy, social, and economic and sustainability risks. The key findings reveal a heterogeneous nature of risk perceptions across different dimensions, with each factor demonstrating varying associations with such perceptions (Table 3).

Multiple linear regression analysis revealed that perceptions of physiological risks were significantly influenced by the establishment of standard medical AI ethics review procedures, years of work experience, familiarity with AI, and training experience, among which the implementation of ethics review protocols demonstrated the most substantial effect. Specifically, working in hospitals with established ethics review procedures was significantly positively associated with physicians’ perception of both misdiagnosis (β=0.355, P=.001) and mistreatment risks (β=0.355, P=.001). Compared with highly experienced physicians, those with no more than 15 years of practice reported higher perception of misdiagnosis risks (β=2.712, P=.02). Physicians unfamiliar with medical AI demonstrated lower perception of misdiagnosis risks than their familiar counterparts (β=–0.235, P=.03). Furthermore, Physicians who had received specialized training showed a significantly more positive perception of misdiagnosis risks compared with those without any training (β=0.230, P=.02).

Regarding the dimension of psychological risk perception, only familiarity with AI demonstrated a statistically significant influence. Other factors may affect the perception of psychological risks through indirect pathways, although their effects are relatively weak and insufficient to achieve statistical significance. Specifically, compared with physicians familiar with AI, those who were very unfamiliar showed a negative correlation in their perception of both technical anxiety among physicians (β=–0.300, P=.007) and psychological anxiety among patients (β=–0.295, P=.006). Similarly, physicians with moderate familiarity also demonstrated a negative correlation in technical anxiety among physicians (β=–0.230, P=.02) and psychological anxiety among patients (β=–0.205, P=.04) compared with those familiar with AI.

In the dimension of data and privacy risk perception, the establishment of medical AI ethics review procedures and training experience demonstrated statistically significant positive effects. Specifically, physicians working in hospitals with established AI ethics review protocols showed higher perception of both data security risks (β=0.230, P=.03) and privacy leakage risks (β=0.245, P=.01) compared with those in institutions without such procedures. Furthermore, individuals who had received specialized training displayed a more acute perception of data and privacy risks than those without training (β=0.220, P=.02).

In the dimension of social risk, hospital level, years of service, familiarity with AI, training experience, and the establishment of medical AI ethics review procedures demonstrated significant driving effects on risk perception, with training experience as the most influential factor. Specifically, physicians who had received specialized training showed higher perception than untrained counterparts in the following dimensions: fairness concerns (β=0.220, P=.03), trust crisis (β=0.265, P=.003), ambiguous accountability (β=0.285, P=.002), and diminished autonomy (β=0.425, P<.001). Compared with highly experienced physicians, those with no more than 15 years of service showed significantly higher perception of trust crisis (β=4.372, P=.008). Physicians unfamiliar with medical AI demonstrated significantly different perceptions in trust crisis (β=–0.260, P=.006) and ambiguous accountability (β=0.155, P=.009). The establishment of medical AI ethics review procedures was associated with significantly higher perception of both trust crisis (β=0.330, P=.001) and ambiguous accountability (β=0.390, P<.001) compared with institutions without such protocols. Notably, physicians in tertiary hospitals showed a negative effect in their perception of occupational impact compared with those in primary hospitals (β=–1.743, P=.02).

In the dimension of economic and sustainability risks, familiarity with AI and training status showed several notable associations. Specifically, physicians’ familiarity with AI showed a statistically significant negative correlation with perceptions of increased financial burden (β=–0.225, P=.04) and resource waste (β=–0.285, P=.01), and a negative correlation that did not reach statistical significance with energy consumption (β=–0.210, P=.06). Physicians who had received specialized training expressed significantly greater concern regarding resource waste (β=0.265, P=.01) and energy consumption (β=0.315, P=.003) compared with those without training.

This study achieved its two aims: (1) constructing a 5-dimensional ethical risk framework for medical AI, and (2) identifying physicians’ perceptions of these risks and identifying key influencing factors via a cross-sectional survey of 600 Chinese physicians. The constructed framework comprises 5 main categories and 15 subcategories. The survey revealed that physicians expressed the highest concern for data and privacy risks, ambiguous accountability, and a physician-patient trust crisis, while economic and sustainability risks received the lowest agreement. Multiple regression analysis further identified that specialized AI training, familiarity with AI, and the establishment of hospital AI ethics review procedures were significant predictors of risk perception.

First, this study breaks through the limitations of a fragmented and isolated approach often found in existing research on medical AI risks, by constructing an ethical risk framework encompassing 5 main categories and 15 subcategories. While conceptual overlaps with existing international frameworks are inevitable given the universal nature of core ethical issues, the distinct contribution of our framework lies in (1) the explicit inclusion of economic and sustainability risks as a core ethical dimension; (2) its empirical derivation from Chinese clinical experts’ lived experiences rather than solely from normative analysis; and (3) its validation through physician cognitive surveys, which provides practical priorities for risk governance. This combined methodological and contextual novelty enhances the framework’s relevance for both local and global medical AI ethics discourse. Compared with current international guidelines, such as the WHO’s “Ethics and Governance of Artificial Intelligence for Health,” which focuses primarily on principled issues like data security, accountability, and fairness [38]; the European Union’s Artificial Intelligence Act, which emphasizes data governance, algorithmic transparency, human oversight, and particularly the explainability of high-risk medical AI systems and patient informed consent [39]; the US Food and Drug Administration (FDA) centers its review of AI medical devices on safety and effectiveness, but systematic ethical assessment has not yet become a mandatory requirement [40]; although relevant white papers in Canada propose professional ethical frameworks, they are mainly limited to single scenarios such as medical imaging and lack a holistic risk perspective across departments and multiple scenarios [41]. Starting from multiple dimensions including physical, psychological, data and privacy, social, and economic and sustainability aspects, this study incorporates numerous specific risks into the framework, such as diagnostic errors, improper treatment, data security, and privacy leakage. Furthermore, it innovatively includes economic and sustainability risks in ethical considerations, which aligns with the goals of “ensuring healthy lives” and “climate action” outlined in the United Nations Sustainable Development Goals [42]. This makes the discussion go beyond simple operations or environmental management. These issues raise fundamental ethical concerns, including equity in health care access, affordability for patients, and intergenerational justice, as rising operational costs may be passed on to users or exacerbate existing disparities in resource allocation across regions and institutions. A systemic, ethics-based examination of such impacts is essential to understand the long-term consequences of medical AI on the health care ecosystem as a whole. This systemic, ethics-based examination of long-term consequences provides a theoretical foundation for the ethical governance of medical AI in China and offers a Chinese approach to global governance.

Our survey revealed that physicians’ ethical concerns are highly concentrated in 3 interconnected dimensions—data and privacy risks, ambiguous accountability, and physician-patient trust crisis. These findings provide evidence-based priorities for AI risk governance.

First, over 67% of physicians expressed strong concern about data security and privacy leakage, aligning with WHO guidelines. Both the European Union’s General Data Protection Regulation (GDPR) and the US HIPAA (Health Insurance Portability and Accountability Act) have established stringent frameworks for medical data protection, highlighting principles such as informed consent, data minimization, and restrictions on cross-border data transfer, thereby providing a legal foundation for data ethics in AI applications. However, practical challenges remain, including the secondary use of data and insufficient algorithmic transparency [43,44]. This high level of concern is particularly salient in the Chinese context; on the one hand, the Personal Information Protection Law is still in its early implementation stage, with 37.5% (225/600) of hospitals lacking standardized AI ethics review procedures; on the other hand, physician-patient relationships in China rely heavily on interpersonal trust, and data leakage risks directly threaten this foundation, thereby amplifying physicians’ professional anxiety. Since the implementation of the Personal Information Protection Law of the People’s Republic of China, significant legislative strengthening has occurred, although enforcement is still in its early stages. The fact that 37.5% (225/600) of hospitals have not yet established standardized AI ethics review procedures reveals gaps in the development of data governance mechanisms within medical institutions. Therefore, there is a practical need to refine detailed rules on data and patient privacy protection, prevent disorderly exploitation and misuse of data in sharing practices, ensure effective informed consent, and translate privacy principles into actionable operational protocols.

Second, 75.7% (454/600) of respondents identified ambiguous accountability as an urgent issue. The European Union’s Artificial Intelligence Act imposes strict regulations on medical AI systems, supplemented by the AI Liability Directive, which facilitates the determination of fault and causation [45]. In 2025, the European Parliament recommended a strict liability system for high-risk AI systems to ensure effective compensation for victims. The United States explores shared responsibility but lacks federal mandatory legislation. WHO guidelines call for human oversight to clarify responsibilities. However, China still lacks clear regulatory documents defining accountability. Moreover, data show that 45.7% (274/600) of physicians are unaware of hospital-level AI ethics procedures. This “fragmented ethical oversight” heightens concerns over uncertain accountability, especially among physicians with low AI literacy, who may avoid using AI for fear of assuming liability.

Third, about 70% (419/600) of physicians agreed that AI weakens emotional communication and threatens patient trust, reflecting disruption of the traditional physician-patient relationship, particularly in China, where interpersonal bonds are highly valued. WHO guidelines note that AI introduces a “physician-patient-AI” model, reducing time and opportunities for face-to-face communication and emotional empathy, which conflicts with China’s communication-centered medical culture. Additionally, WHO guidelines emphasize that medical AI is characterized by “technical black boxes” and “explanation black boxes.” The low transparency of algorithmic decision-making makes it difficult for patients to comprehend the logic behind AI-generated recommendations. This “uncertainty” can amplify patients’ skepticism toward diagnostic results, triggering a crisis of trust. The high level of concern among Chinese physicians regarding such risks stems from their acute awareness of the importance of “interpersonal bonds” and the “vulnerability of trust” within local physician-patient relationships. The trust crisis is influenced by specialized training, AI familiarity, and institutional ethics review standards. Therefore, AI ethics training and enhanced AI literacy are essential to rebuilding trust.

Finally, concerns about fairness and occupational impact varied significantly by hospital level. In our sample, 62.5% (375/600) of physicians worked in tertiary hospitals, and only 21% (126/600) had received specialized AI training, potentially underestimating the concerns of primary care physicians who face greater resource constraints. For instance, agreement rates for fairness (54.7%) and digital divide (63.3%) may be underestimated due to the limited representation of primary care physicians, who face greater resource constraints. Similarly, low concern about occupational impact likely reflects overrepresentation of tertiary hospital physicians, who feel less threatened by AI substitution. Thus, current findings may reflect a technology-advantaged perspective; actual risk perception may be higher, especially in resource-limited settings. Similarly, the relatively low physician agreement on economic and sustainability risks likely reflects clinicians’ primary focus on immediate clinical responsibilities, diagnostic safety, treatment effectiveness, and patient communication rather than a lack of ethical sensitivity. Physicians have limited bandwidth to attend to broader sustainability issues, while critical from a policy and global governance perspective, remain peripheral to daily medical decision-making. The FDA encourages the use of an “algorithmic bias evaluation” framework to identify and mitigate discrimination based on race, gender, or geographic region [46,47]; while the United Kingdom’s National Institute for Health and Care Excellence recommends addressing job replacement anxiety through role transformation [48]. In China, primary care institutions lack technology access, training resources, and institutional support, making fairness and professional transition issues more pronounced. This study found that over half of the respondents expressed concerns about diagnostic and treatment inequities caused by flaws in AI algorithms, as well as the digital divide resulting from uneven adoption of AI technologies. Although most physicians were not overly worried about AI threatening their careers, primary care providers reported significantly higher concern than tertiary hospital peers. This disparity reflects the uneven distribution of medical resources and career opportunities across different levels of health care institutions. Therefore, a dual approach is needed—promoting AI integration into primary care to reduce the digital divide, while strengthening ethical governance in tertiary hospitals, clarifying the auxiliary role of AI, and alleviating anxiety.

This study has several limitations. First, the expert sample was relatively small and predominantly clinical, limiting the exploration of legal and digital risk dimensions. Second, convenience sampling overrepresented physicians from tertiary hospitals and senior titles, reducing generalizability to primary care and less-experienced clinicians. Third, the lack of patient perspectives may overlook key ethical risks, such as autonomy, informed consent, and trust. Fourth, the low agreement on economic and sustainability risks may reflect measurement limitations, as items did not explicitly link these risks to physicians’ clinical responsibilities. Future research should diversify expert samples, include patient views, use stratified sampling, and develop more context-sensitive items to assess sustainability risks.

This study systematically examines the ethical risks of medical AI in China and their impact on physicians’ perceptions. The findings carry several broader implications for policy, governance, and practice. The main conclusions are as follows:

A multidimensional ethical risk framework for medical AI was constructed. Using grounded theory, we constructed a framework with five main categories: (1) physiological risks, (2) psychological risks, (3) data and privacy risks, (4) social risks, as well as (5) economic and sustainability risks—along with 15 subcategories. Despite low physician agreement on economic or sustainability risks, its inclusion is theoretically and ethically justified for sustainable health care. This framework overcomes fragmented existing research and provides a foundation for systematic risk governance.

Physicians’ perceptions and primary concerns regarding the ethical risks of medical AI were revealed. A survey of 600 physicians revealed concentrated concerns, such as data privacy leakage (over 67% expressed a high level of concern), ambiguous accountability (75.7% believed it requires urgent attention), and a physician-patient trust crisis (approximately 70% indicated concern). Thus, data security, accountability, and preserving physician-patient relationships are foremost ethical priorities.

Multiple regression showed that physicians’ familiarity with AI, specialized AI training, and hospital-level AI ethics review procedures significantly affect risk awareness. Specialized training increased sensitivity to data privacy risks and social risks, such as fairness concerns, trust crisis, and ambiguous accountability. Established ethics review protocols positively correlated with awareness of physiological risks, data and privacy risks, and social risks.

In summary, this study enriches the theoretical framework and provides empirical evidence for targeted governance. Future efforts should strengthen AI ethics training for physicians, improve institutional ethics review, clarify accountability systems, and integrate AI into primary care to reduce the digital divide.