Medical students perceptions and attitudes toward the use of generative artificial intelligence in clinical decision-making: a nationwide cross-sectional survey in China

Abstract

Background

The deep integration of artificial intelligence (AI) into healthcare is reshaping medical practice and education globally. As an emerging technology, generative AI (GenAI) demonstrates significant potential for application in clinical decision-making. Systematically understanding medical students’ perceptions and attitudes toward GenAI is crucial for promoting its responsible implementation in the medical field.

Objective

This study aimed to investigate Chinese medical students’ perceptions, usage behaviors, and attitudes regarding the use of GenAI for clinical decision-making.

Methods

This exploratory cross-sectional study was conducted via an online questionnaire from January to March 2025. A total of 1062 medical students from 168 universities and colleges across 29 provinces in China were recruited through convenience sampling. The survey, developed based on the Technology Acceptance Model (TAM) and validated by expert review and pilot testing, descriptively assessed three dimensions: usage, perceptions, and attitudes toward GenAI in clinical decision-making. Descriptive statistics, including frequencies and 95% confidence intervals, were used for data analysis.

Results

The vast majority of students (99.4%, n = 1056) reported prior experience with GenAI. The primary application was course learning (71.8%, 95% CI [0.690–0.744]); in contrast, direct use in clinical decision-making was reported less frequently (44.0%, 95% CI [0.410–0.470]). Students widely recognized GenAI’s benefits in broadening knowledge (73.4%, 95% CI [0.706–0.759]), fostering multi-perspective clinical thinking (67.8%, 95% CI [0.649–0.705]), and improving efficiency (63.1%, 95% CI [0.601–0.659]). They also noted significant limitations: primarily its inability to account for individual patient differences in diagnosis (70.7%, 95% CI [0.679–0.734]) and susceptibility to input data bias (65.6%, 95% CI [0.627–0.684]). Most students (71.7%, n = 762) were willing to use GenAI in the future, yet strongly opposed its complete replacement of healthcare professionals (79.4%, n = 843) and advocated for safeguards such as strict output auditing (69.6%, 95% CI [0.668–0.723]).

Conclusion

This study reveals that medical students maintain a “cautious embrace” attitude toward GenAI: actively utilizing them while consistently emphasizing the central importance of professional judgment. This finding suggests that medical education should focus on cultivating future healthcare professionals who can skillfully employ GenAI as a supportive tool, while steadfastly adhering to critical AI literacy.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12909-026-08698-7.

Introduction

The rapid development of Artificial Intelligence (AI) is profoundly changing the practice model of healthcare worldwide [1]. As a significant technological advancement, Generative Artificial Intelligence (GenAI), with its powerful data learning and content generation capabilities, has demonstrated extensive application potential in multiple fields. In the clinical field, GenAI is increasingly applied to assist in formulating treatment plans, analyzing medical images, optimizing nursing workflows, predicting health risks, and supporting public health responses. In non-clinical fields, it also plays an increasingly important role in medical education, patient communication, and operational management [2–5]. As the core stage for training future healthcare professionals, medical education is undergoig a critical transformational period integrating with AI [6]. Guided by the Technology Acceptance Model (TAM), which posits that technology adoption is driven by the balance between perceived usefulness and perceived risks/ease of use, this study examines how these core constructs shape medical students’ attitudes and behavioral intentions toward using GenAI in the high-stakes context of clinical decision-making [7].

Investigations from multiple global regions indicate that medical students generally hold a cautiously optimistic attitude toward AI, but there are significant differences in their specific acceptance levels and application methods. For instance, Swedish medical students have the highest acceptance of AI (71%) and 64% of them actively use GenAI to assist their clinical learning, indicating strong perceived usefulness in driving medical progress [8]. In contrast, while Australian medical students recognize AI’s positive impact (74.4%), only 44.7% are willing to use it, with perceived risks centered on weakening doctor-patient empathy [9]. Pakistani medical students, report positive attitudes toward AI (80.3%) but significant concern about job replacement (44.8%), highlighting another salient perceived risk [10]. Collectively, these international findings underscore that medical students’ acceptance is not a simple binary but a complex calculus shaped by the TAM’s core constructs. While recognition of AI’s perceived usefulness is widespread, its adoption is critically moderated by context-laden perceived risks, such as fears over eroded patient empathy or professional displacement. Furthermore, such concerns are not confined to students; similar worries about liability, workflow changes, and ethical implications have been documented among clinical staff [11]. However, the applicability of these specific risk-and-usefulness profiles to the Chinese context remains an open question. In China, AI in healthcare is being rapidly propelled by distinctive national strategic policies, aiming to establish standards and promote implementation in areas like assisted diagnosis [12]. This creates a unique “top-down” promotion context alongside specific socio-cultural dynamics in clinical practice. Therefore, directly extrapolating findings from other settings is insufficient; a focused investigation within China’s unique policy and cultural landscape is imperative to understand the local configuration of perceived usefulness and perceived risks among future practitioners.

The potential of GenAI extends into medical education, where it is explored for virtual patient simulation, content generation, and adaptive learning [13–15]. However, the ultimate goal of medical education is to cultivate competent clinical decision-makers. A critical, yet understudied, gap exists at the nexus of educational tool adoption and core clinical competency development. Specifically, there is scant research on how medical students, who are in the process of forming their professional judgment, perceive, trust, and intend to integrate GenAI into the very heart of clinical work: making decisions about diagnosis, treatment, and prognosis. This gap is especially pronounced in China, where rapid policy promotion contrasts with a scarcity of large-scale, empirical evidence on the readiness of medical students, the frontline future practitioners, to employ GenAI for specific clinical tasks.

Based on the background, this study aims to investigate medical students’ perceptions and attitudes toward using GenAI for clinical decision-making in China. For this study, clinical decision-making is operationally defined as AI-assisted support in key diagnostic-therapeutic tasks, including differential diagnosis generation, treatment recommendation, and disease progression prediction. Therefore, by collecting large sample data from multiple medical universities and colleges across China, this study aims to describe the current perceptions, attitudes, and behavioral intentions of Chinese medical students regarding the use of GenAI for clinical decision-making, and to explore how the core constructs of the TAM (perceived usefulness, perceived risks) manifest within this specific population and context.

Method

Study design

This cross-sectional survey aimed to investigate the perceptions, attitudes and willingness to use GenAI for clinical decision-making among Chinese medical students. Data were collected online from January to March 2025.

Participants and eligibility

The study recruited medical students currently enrolled in undergraduate or postgraduate programs in China. All eligible students were invited to participate, with no restrictions based on age, gender, or year of study. A survey was conducted among 1062 medical students from 168 universities and colleges in 29 provinces in China.

Data collection and procedures

A convenience sampling approach was used without active stratification or quotas for province, discipline, or institution type. The survey was disseminated through an open link. This study used Questionnaire Star (Changsha Ranxing Information Technology Co., LTD, China) to create the survey questionnaire and generate a two-dimensional code for the questionnaire link. At the beginning of the questionnaire, the research purpose, principle of voluntary participation, anonymity, and data confidentiality were clearly stated. All participants provided informed consent before proceeding. Participants accessed and completed the questionnaire by scanning the QR code, and all data were collected directly through the Questionnaire Star system.

Participants were mainly recruited through medical-related online communities (primarily professional medical education and student forums such as MedLive, DXY.cn, WeChat and QQ) and campus channels (student associations, academic affairs offices, course coordinators and university learning management systems, such as ChaoXing, Rain Classroom).

A total of 1070 submissions were received. After collection, all submissions were screened. Cases flagged with identical IP addresses were manually reviewed for patterns suggesting duplicate entries (e.g., identical response sets within a short timeframe) before exclusion. After rigorous data cleaning, including eliminating duplicate IP addresses, obvious logical errors, or any missing responses in the core analytical sections, 8 invalid questionnaires were excluded. Eventually, 1062 valid questionnaires were obtained. Consequently, the final dataset of 1062 responses contained complete data for all core items related to perceptions and attitudes, as these questions were set as mandatory in the survey.

Survey instrument: development and validation

Questionnaire design and content validity

The design of this survey instrument was guided by an exploratory-descriptive research objective: to capture a wide spectrum of specific attitudes and behaviors rather than to measure a few latent psychological constructs. Therefore, the questionnaire was conceptualized as a comprehensive item set, not as a summated rating scale. Items were generated based on the theoretical framework of the TAM and a review of pertinent literature, organized thematically around cognitive, technical, and institutional dimensions.

To ensure content validity, relevance, and clarity, the initial item set was reviewed by an expert panel comprising two medical education specialists and one clinician with AI research experience. Through informal consultation, the experts provided feedback on item wording, comprehensiveness, and appropriateness for the target population. Their suggestions were used to refine several items, thereby enhancing the content validity for our specific context.

Pilot testing and instrument refinement

A pilot test was conducted with 50 medical students who met the study’s inclusion criteria. The primary goals of the pilot were practical and qualitative: to estimate average completion time, identify ambiguous or difficult-to-understand items, and assess the logical flow of the questionnaire. Participants provided open-ended feedback, which led to minor wording adjustments to improve clarity and precision. This process is critical for ensuring the feasibility and appropriateness of a survey tool in descriptive research.

Final structure and transparency

The final questionnaire consisted of 20 items across four sections: basic information (6 items), awareness and usage of GenAI (4 items), perceptions of GenAI for clinical decision-making (5 items) and attitudes of GenAI for clinical decision-making (5 items). The complete questionnaire is provided in Supplementary File 1 to ensure full transparency.

Approach to reliability

Given that the questionnaire is an item set where each question measures a distinct, standalone perception or behavior, calculating a traditional Cronbach’s alpha for the entire tool or for theoretical subgroups would be methodologically inappropriate. Internal consistency metrics are suitable when multiple items are designed to be inter-correlated and summed to represent a single latent variable, which was not the aim of this study.

Consequently, all primary analyses in this study were performed at the individual item level. We report descriptive statistics (frequencies and percentages) for each item separately. The “dimensions” discussed in the results (e.g., cognitive, technical) are post hoc, thematic groupings used for narrative organization of findings from related individual items, not statistically derived or validated constructs.

Data analysis

Data were independently entered and cross-checked by two researchers. The statistical analysis was performed with the SPSS 27.0 (IBM Corp, Armonk, NY, USA). Consistent with the study’s exploratory and descriptive aims, analyses were restricted to descriptive statistics. Categorical variables are presented as frequencies and percentages. For key proportional outcomes, 95% confidence intervals were calculated using the Wilson score interval method to indicate the precision of the estimates. The continuous variable (age) is summarized as mean and standard deviation. No inferential statistical tests (e.g., hypothesis testing, regression) were performed, as the non-probability sampling method does not support population-level generalizations or causal inference.

Result

Demographic data

This study ultimately included 1062 medical students from 168 universities and colleges across 29 provinces in China. As summarized in Table 1, the sample spanned all undergraduate and postgraduate levels as well as all 11 primary medical disciplines. Participants had a mean age of 20.83 ± 1.96 years (range 17–30), and the majority were female (66.5%, n = 706) and undergraduates in their first to fourth years (92.5%, n = 983). The most represented majors were nursing (25.1%, n = 267), clinical medicine (19.7%, n = 209), and basic medicine (12.3%, n = 131). Over half (51.1%, n = 543) reported 3–5 h of daily study time.

Table 1.

Demographic characteristics of participants (N = 1062) (N, %)

Characteristics	N	%
Sex
Male	356	33.5%
Female	706	66.5%
Grade level
Year 1 undergraduate	182	17.1%
Year 2 undergraduate	435	41.0%
Year 3 undergraduate	151	14.2%
Year 4 undergraduate	215	20.2%
Year 5 undergraduate a	40	3.8%
Master’s/Doctoral Candidate	39	3.7%
Majors b
Basic Medicine	131	12.3%
Clinical Medicine	209	19.7%
Stomatology	87	8.2%
Public Health and Preventive Medicine	75	7.1%
Chinese Medicine	62	5.8%
Chinese and Western Integrative Medicine	51	4.8%
Pharmaceutical Science	73	6.9%
Science of Chinese Pharmacology	46	4.4%
Traditional Chinese Medicine Forensic medicine	29	2.7%
Medical technology	32	3.0%
Nursing	267	25.1%
The time you spend each day on learning medical knowledge
2 h or less	250	23.5%
3–5 h	543	51.1%
6–10 h	229	21.6%
11 h or more	40	3.8%

^a Year 5 undergraduate: refers to students in the final year of certain programs, typically including medicine and related professional degrees in China’s five-year undergraduate system

^b Majors: the item was a single-choice question. Participants selected their one primary field of study; therefore, the percentages sum to 100%

Usage of GenAI

Medical students reported high familiarity and extensive engagement with GenAI (Table 2). The overwhelming majority (99.4%, n = 1056) had prior experience with GenAI, and among users, regular usage was common, with 64.3% (n = 683) employing it several times per week. The most frequently used platforms were Doubao, ChatGPT, and Wenxin Yiyan.

Table 2.

Usage of GenAI (N = 1062)

Questions	N (%)	95% CI
1. Do you know about GenAI?
Very familiar	112 (10.5%)	0.088–0.125
Relatively familiar	466 (43.9%)	0.409–0.469
Generally familiar	318 (29.9%)	0.273–0.328
Slightly familiar	141 (13.3%)	0.114–0.155
Completely unfamiliar	25 (2.4%)	0.016–0.035
2. How often do you use GenAI?
Every day	167 (15.7%)	0.137–0.180
Several times a week	516 (48.6%)	0.456–0.516
Several times a month	164 (15.4%)	0.134–0.177
Occasionally, with no fixed frequency	209 (19.7%)	0.174–0.222
Never	6 (0.6%)	0.003–0.012
3. Which GenAI tools do you usually use? a
DeepSeek	402 (37.9%)	0.350–0.408
ChatGPT	605 (57.0%)	0.540–0.599
Wenxin Yiyan	510 (48.0%)	0.450–0.510
Tongyi Qianwen	215 (20.2%)	0.179–0.228
iFlytek Spark	316 (29.8%)	0.271–0.326
Kimi	423 (39.8%)	0.369–0.428
Doubao	627 (59.0%)	0.561–0.620
Others c	9 (0.8%)	0.004–0.016
4. In which areas do you primarily use GenAI? a
Course learning (such as assisting in understanding course content, completing assignments, etc.)	762 (71.8%)	0.690–0.744
Research work (such as literature review, data analysis, etc.)	613 (57.7%)	0.547–0.607
Clinical practice (such as clinical decision-making, providing diagnosis suggestions, etc.)	467 (44.0%)	0.410–0.470
Information searching	694 (65.3%)	0.624–0.682
Answering questions	634 (59.7%)	0.567–0.626
Conversational chatting	339 (31.9%)	0.292–0.348
Others b	5 (0.5%)	0.002–0.011

^a Multiple-choice question (Items 3–4). For multiple-choice questions, percentages represented the proportion of respondents selecting each option; percentages sum to > 100%

^b Other responses consisted of “No” or no additional specification

^c Other responses (n = 10): The participants wrote “Xiao Ai, Zhipu Qingyan, Tianggong, Quark AI, Hailuo AI, Baidu Assistant, sider”

Usage was heavily oriented toward academic and learning support. The primary applications included course learning (71.8%, 95% CI [0.690–0.744]), information searching (65.3%, 95% CI [0.624–0.682]), answering questions (59.7%, 95% CI [0.567–0.626]), and research work (57.7%, 95% CI [0.547–0.607]). In contrast, direct application to clinical practice tasks, such as clinical decision-making or diagnostic support, was notably less prevalent, reported by only 44.0% (95% CI [0.410–0.470]) of respondents.

Perceptions of using GenAI in clinical decision-making

Medical students’ perceptions of GenAI in clinical decision-making revealed a balanced recognition of its utility and salient limitations (Table 3). Their views on its potential value were multifaceted. A majority acknowledged its core advantages in broadening knowledge coverage (73.4%, 95% CI [0.706–0.759]), fostering multi-perspective clinical thinking (67.8%, 95% CI [0.649–0.705]), and improving efficiency (63.1%, 95% CI [0.601–0.659]). Students primarily valued its application in predicting disease progression (23.5%, 95% CI [0.211–0.262]), recommending treatments (22.4%, 95% CI [0.200–0.250]), and assisting diagnosis (20.5%, 95% CI [0.182–0.231]).

Table 3.

Perceptions of using GenAI for clinical decision-making (N = 1062)

Questions	N (%)	95% CI
1. How accurate do you think GenAI for clinical decision- making?
Very accurate, generally matches the actual situation	82 (7.7%)	0.063–0.095
Mostly accurate, provides useful reference value	502 (47.3%)	0.443–0.503
Moderately accurate, needs to be combined with other sources for judgment	389 (36.6%)	0.338–0.396
Poor accuracy, limited reference value	77 (7.3%)	0.058–0.090
Completely inaccurate; highly misleading	12 (1.1%)	0.006–0.020
2. When using GenAI for clinical decision-making, which aspect are you most interested in?
Disease diagnosis suggestions	218 (20.5%)	0.182–0.231
Treatment plan recommendations	238 (22.4%)	0.200–0.250
Prediction of disease progression trends	250 (23.5%)	0.211–0.262
Interpretation of examination results	196 (18.5%)	0.162–0.209
Patient care recommendations	143 (13.5%)	0.115–0.157
Others b	17 (1.6%)	0.010–0.025
3. What roles do you think GenAI can play for clinical decision-making? a
Quickly provide possible directions for disease diagnosis	714 (67.2%)	0.644-0.700
Assist in analyzing the correlations between patients’ symptoms and medical history	783 (73.7%)	0.710–0.763
Generate detailed disease analysis reports	621 (58.5%)	0.555–0.614
Help in formulating treatment plan	578 (54.4%)	0.514–0.574
Others c	5 (0.5%)	0.002–0.011
4. What do you think to be the advantages of using GenAI for clinical decision-making? a
Rapid provision of condition-related information to save time	670 (63.1%)	0.601–0.659
Broad knowledge coverage to complement one’s own knowledge gaps	779 (73.4%)	0.706–0.759
Analyze conditions from multiple perspectives to broaden clinical thinking	720 (67.8%)	0.649–0.705
Clear and understandable expression to understand	441 (41.5%)	0.386–0.445
Others b	3 (0.3%)	0.001–0.008
5. What do you think to be the limitations of using GenAI for clinical decision-making? a
Lack of clinical practical experience, resulting in less practical recommendations	608 (57.3%)	0.543–0.602
Cannot accurately account for individual patient differences	751 (70.7%)	0.679–0.734
Information may not be timely updated, potentially providing outdated knowledge	557 (52.4%)	0.494–0.554
Susceptible to input information bias, leading to erroneous or misleading results	697 (65.6%)	0.627–0.684
Reliance on large datasets, where data quality significantly impacts analysis outcomes	495 (46.6%)	0.436–0.496
Others b	2 (0.2%)	0.001–0.007

^a Multiple-choice question (Items 3–5). For multiple-choice questions, percentages represented the proportion of respondents selecting each option; percentages sum to > 100%

^b Other responses consisted of “No” or no additional specification

^c Other responses (n = 5): One participant wrote “searching for literature and gathering information”. The rest are all wrote “No”

However, pronounced concerns tempered this recognition of utility. The most frequently cited limitation was its inability to account for individual patient differences in diagnosis (70.7%, 95% CI [0.679–0.734]). Students also highlighted its vulnerability to biases in input data (65.6%, 95% CI [0.627–0.684]). This cautious outlook was further reflected in the fact that over one-third (36.6%, 95% CI [0.338–0.396]) explicitly emphasized the necessity of complementing AI outputs with independent professional judgment, while just over half (55.0%, n = 584) considered its outputs to be at least “mostly accurate”.

Attitudes toward using GenAI in clinical decision-making

Medical students expressed a pragmatic yet cautious stance toward GenAI in clinical decision-making (Table 4). While a high proportion (71.7%, n = 762) reported willingness to use it in the future, the vast majority (79.4%, n = 843) opposed the notion of it completely replacing healthcare professionals.

Table 4.

Attitudes toward using GenAI in clinical decision-making (N = 1062)

Questions	N (%)	95% CI
1. Do you think GenAI can replace healthcare professionals in clinical decision-making?
Completely agree	107 (10.1%)	0.084–0.120
Partially agree (it can replace professionals in certain specific fields)	354 (33.3%)	0.306–0.362
Serve only as an auxiliary tool, cannot fully replace professionals	489 (46.1%)	0.431–0.491
Completely disagree	83 (7.8%)	0.063–0.096
Uncertain	29 (2.7%)	0.019–0.039
2. What do you think is the most important principle that medical students should follow when using GenAI for clinical decision-making during their study period?
Ensure the authenticity and accuracy of the input information	172 (16.2%)	0.141–0.185
Make judgments based on professional knowledge and avoid blind reliance on AI suggestions	557 (52.4%)	0.494–0.554
Protect patient privacy and prevent information leakage	211 (19.9%)	0.176–0.224
Scrutinize the reliability of the AI-generated information	101 (9.5%)	0.079–0.114
Others b	21 (2.0%)	0.013–0.030
3. Are you willing to use GenAI for clinical decision-making in future studies and work?
Very willing, will actively use it	165 (15.5%)	0.135–0.178
Relatively willing, depend on the situation	597 (56.2%)	0.532–0.592
Neutral, would follow others’ practices	192 (18.1%)	0.159–0.205
Not very willing, worry about risks and responsibilities	92 (8.7%)	0.071–0.105
Completely unwilling, do not trust its reliability	16 (1.5%)	0.009–0.024
4. What improvements or optimizations do you expect GenAI to make in clinical decision-making? a
Improve the accuracy and practicality of the information	634 (59.7%)	0.567–0.626
Enhance the consideration of individual patient differences	773 (72.8%)	0.700-0.754
Strengthen interactivity with users to better understand their needs	645 (60.7%)	0.578–0.636
Update the knowledge bases to reflect the latest medical advances	544 (51.2%)	0.482–0.542
Others c	6 (0.6%)	0.003–0.012
5. What do you think are the essential safeguard mechanisms that medical institutions must establish? a
Clarify the attribution of medical accident liability	637 (60.0%)	0.570–0.629
Strictly auditing the outputs of GenAI	739 (69.6%)	0.668–0.723
Improving the protection of patient privacy and data security	693 (65.3%)	0.623–0.681
Avoid excessive reliance and maintain one’s own clinical thinking and judgment ability. Avoiding over-reliance to maintain the autonomy of clinical decision-making	669 (63.0%)	0.601–0.659
Continuously update the medical knowledge and algorithms of GenAI	501 (47.2%)	0.442–0.502
Others b	3 (0.3%)	0.001–0.008

^a Multiple-choice question (Items 4–5). For multiple-choice questions, percentages represented the proportion of respondents selecting each option; percentages sum to > 100%

^b Other responses consisted of “No” or no additional specification

^c Other responses (n = 6): One participant wrote “The data varies significantly in different regions, so thorough data research is necessary”. Other participants wrote “No”

This caution was reflected in their specific demands. Students emphasized the need for technological refinement, particularly in personalized diagnosis and treatment (72.8%, 95% CI [0.700-0.754]). Concurrently, they strongly advocated for robust safeguard mechanisms, including strict output auditing (69.6%, 95% CI [0.668–0.723]) and enhanced data privacy protection (65.3%, 95% CI [0.623–0.681]).

Crucially, this pragmatic stance was underpinned by a prioritization of professional oversight. Most medical students (63.0%, 95% CI [0.601–0.659]) cautioned against over-reliance on AI, and when asked to identify the single most important principle for use during training, over half (52.4%, 95% CI [0.494–0.554]) endorsed “making judgments based on professional knowledge and avoiding blind reliance on AI suggestions”. Concerns regarding liability attribution were also prominent (60.0%, 95% CI [0.570–0.629]).

Discussion

This nationwide survey reveals that Chinese medical students exhibit a stance of “cautious embrace” toward GenAI in clinical decision-making. Characterized by near-universal adoption (99.4%), substantial willingness for future use (71.7%), yet profound resistance to its role as a replacement for healthcare professionals (79.4%), this attitude reflects a nuanced and pragmatic calculation. Students actively harness GenAI’s utility for learning while simultaneously demanding robust safeguards, emphasizing professional oversight, and highlighting its technical limitations in personalization. This discussion interprets these findings through cognitive, technical, and institutional dimensions, exploring the questions they raise and identifying potential directions for medical education in the era of GenAI.

Cognitive dimension: from tool dependence to cultivating critical AI literacy

The data reveal a dual attitude among Chinese medical students toward GenAI, which is well-explained by the TAM. The high adoption rate (99.4%) and frequent use, primarily for academic efficiency, underscore a strong perceived usefulness rooted in GenAI’s capacity as an information-processing and learning-augmentation tool. This is similar to medical students in Sweden, who also use AI actively for learning [8].

However, the Chinese cohort distinguishes itself through the intensity of its perceived risks. While students in places like Pakistan may worry about losing jobs to AI [10], the students in China are more concerned about how it might affect the core of being healthcare professionals. Most reject the idea of AI replacement (79.4%). At the same time, many stress the need to avoid over-reliance (63.0%) and to make independent professional judgments (52.4%). Together, this shows their main worry is preserving clinical autonomy, diagnostic responsibility, and the integrity of the doctor-patient relationship. They are practical: they will use helpful technology, but not at the expense of what makes medicine human.

The students’ pronounced wariness toward over-reliance can be understood as a preemptive guard against cognitive offloading [16], a process where delegation of analytical tasks to AI could, over time, undermine the development and maintenance of their own clinical reasoning [17, 18]. Their “use-but-verify” approach is a conscious buffer against this risk.

Therefore, the results of this study suggest that medical education needs to consider how to cultivate students’ critical skills in using GenAI [19]. This is not merely about teaching technical operations, but also about fostering a mindset that enables future healthcare professionals to: critically evaluate the output of GenAI; understand its inherent limitations (such as algorithmic biases, lack of contextual considerations [20]); actively define the reasonable boundaries of GenAI assistance.

This study, as a cross-sectional survey, cannot provide a definitive answer, but it does point out a key issue. It reminds us that in the era of GenAI, a core goal of medical education might not be to cultivate passive users of tools, but to train healthcare professionals who can lead human-AI collaboration. The honed clinical judgment will always guide and supervise the integration and application of GenAI.

Technological dimension: challenges and evolution from generic algorithms to individualized care

Medical students provided a clinically astute critique of GenAI’s capabilities. While most acknowledged its utility in finding data correlations (73.7%), a majority pinpointed its core failing: the inability to account for individual patient differences (70.7%). This directly fueled the strongest demand for improved personalization (72.8%). This finding suggests that students are not just evaluating a tool’s features, but assessing its fitness for a fundamental principle of medicine: care is inherently contextual and personalized.

This “personalization gap” highlighted by students reflects a foundational principle of clinical reasoning they are internalizing. Their widespread concern about this limitation indicates that the unique circumstances of individual patients are more important than the general output of the algorithm. Consequently, their demand for better personalization can be interpreted as a call for GenAI to evolve from a generic information tool into a context-aware clinical partner.

These exploratory findings raise a thought-provoking question for medical education. Since students have already recognized this crucial limitation of GenAI, the next reasonable step in curriculum design might be to focus on developing their ability to manage this limitation. For instance, in teaching, some complex cases that current GenAI might be unable to handle could be simulated, such as cases involving subtle psychological and social factors [21, 22], rare disease manifestations [18, 23], or specific cultural backgrounds [18, 24]. The purpose of this training is to consciously strengthen students’ own clinical judgment, making it the key to compensating for the shortcomings of GenAI.

It must be noted that, as a cross-sectional study, these findings are not intended to prescribe a specific curriculum, but rather to highlight an area of education that requires focused research and development in the future. From this perspective, the technical deficiencies pointed out by the students have actually helped us clearly understand which continuously evolving human professional skills remain indispensable in the future GenAI-assisted clinical environment.

Institutional dimension: liability framework in the era of algorithmic assistance

The data reveal a tension in how medical students view GenAI: while most believe it can improve diagnostic efficiency (67.2%), a majority simultaneously demand clearer liability definitions (60.0%) and insist on strict output review (69.6%). This suggests that while students are looking forward to the convenience brought by GenAI, they are also very concerned about the accompanying responsibility issues.

This concern is likely to stem from the characteristics of GenAI’s application in the actual medical environment. The decision-making process of GenAI is usually opaque, while clinical work requires quick judgments. In such circumstances, once a problem arises, it is difficult to determine whether the responsibility lies with the healthcare professionals who uses the GenAI or with the technicians who developed the GenAI. The students’ concerns precisely reflect their anticipation of this potential predicament they might face in the future professional environment.

Therefore, the results of this study suggest that when promoting GenAI medical applications, it may be necessary to simultaneously establish a corresponding responsibility framework. One possible approach is to differentiate responsibilities based on the risk level of the GenAI-assisted tasks. For example, for low-risk tasks such as organizing medical records, doctors can simply conduct routine confirmations. For high-risk tasks such as assisting in diagnosis and formulating treatment plans, a more rigorous process is required, including documenting the basis for decisions and requiring the technology provider to provide explanations of the algorithm logic, to ensure that the entire process is traceable and reviewable.

It is important to note that this study is a cross-sectional survey, which reveals the current concerns and expectations of the student group, but it cannot predict how these attitudes will affect future clinical behaviors or decisions. The significance of these findings lies in highlighting a key issue: for GenAI to be truly trusted and integrated into healthcare, in addition to technological progress, establishing clear and reasonable liability systems may be equally crucial.

Implications for higher medical education

Integrating GenAI into clinical practice presents both opportunities and challenges for medical education. Based on the findings of this study, we attempt to propose several directions for higher education institutions to prepare future healthcare professionals for human-AI collaboration. As curriculum quality and faculty support are established drivers of student outcomes in health professions education, GenAI should be integrated within these broader institutional support structures rather than treated as a standalone tool [25].

Cultivating critical AI literacy in the course

Our finding that over half (52.4%) of medical students prioritized “making judgments based on professional knowledge” as the core principle for using GenAI suggests a clear pedagogical need. In response, medical education could shift its focus from merely teaching functional tool use to systematically fostering critical GenAI literacy, centered on developing supervisory and intervention capabilities.

Concretely, curricula could be designed to help students learn to monitor GenAI outputs, critically assess their validity and limitations (e.g., algorithmic bias, lack of context), and confidently exercise professional judgment to decide whether to accept or modify GenAI suggestions. This could be implemented through specific teaching methods, such as case studies analyzing GenAI diagnostic errors or simulation exercises designed to train students in recognizing and responding to algorithmic bias.

Through these practical pedagogical approaches, students can gradually build the necessary vigilant mindset and skills to effectively supervise GenAI systems. This would enable GenAI to serve as a competent assistant in clinical work, while ensuring that healthcare professionals retain ultimate responsibility for patient safety and clinical outcomes.

Building an integrated AI curriculum across disciplines

In response to the core concerns in our survey, such as the significant limitations of personalized GenAI and the strong demand for clear accountability, integrating GenAI literacy systematically into existing medical curricula seems to be a more constructive approach than treating it as an independent discipline.

Specifically, this integration could happen in two directions. Vertically, core GenAI concepts could be introduced during foundational stage and further explored and applied in clinical training, particularly in complex decision-making scenarios. Horizontally, relevant competencies could be mapped to specific disciplines: pathology training might include exercises in validating GenAI-assisted outputs; clinical medicine seminars might use GenAI-generated differential diagnoses for critical appraisal; and ethics modules could address topics like algorithmic bias and accountability.

A key enabler for this approach is supporting faculty development. Equipping educators with the necessary knowledge and teaching skills is essential to guide students effectively within this new paradigm, helping them become discerning partners who use GenAI to enhance clinical care.

Co-designing instruction for human-AI collaboration

Given students’ emphasis on the need for oversight and independent judgment, a promising instructional shift could involve thoughtfully integrating GenAI as a collaborative partner in the learning process, with the educator guiding its use.

For example, instructors might employ GenAI-powered adaptive virtual patients that present evolving case scenarios. Students could engage in an iterative dialogue with the GenAI, critiquing its diagnostic suggestions and justifying their own clinical reasoning. The educator’s role would then focus on facilitating debriefings that examine the GenAI’s logic, discuss its potential limitations or biases, and reinforce students’ accountability for the final clinical decision.

To support this shift, assessment methods could also be adapted. Rubrics might begin to evaluate how students interrogate GenAI-generated content, manage disagreements with its recommendations, and ultimately demonstrate independent clinical judgment. This approach aims to ensure that technology becomes a tool for strengthening essential human skills, such as higher-order reasoning, ethical discernment, and empathetic communication, rather than replacing them.

Conclusion and outlook

This exploratory descriptive study reveals a stance of “cautious embrace” toward GenAI among Chinese medical students in clinical decision-making, characterized by high adoption coupled with a strong desire to preserve clinical autonomy and judgment. These findings suggest a corresponding need in medical education: to cultivate students’ ability to work wisely with GenAI. This involves fostering critical AI literacy for output evaluation, strengthening independent judgment in complex scenarios, and cultivating an awareness of attendant ethical and professional responsibilities. Therefore, incorporating these competencies into medical education is recommended. By doing so, we could educate future healthcare professionals who are both skilled in using GenAI and committed to the patient-centered values of healthcare. This would help ensure that GenAI serves as a true aid to medical practice, supporting rather than replacing the role of the healthcare professionals.

Limitations and future researches

This study has several limitations that should be considered when interpreting its findings, which are inherently exploratory and descriptive in nature. First, the cross-sectional design captures attitudes at a single point in time. It cannot establish causality or track how perceptions evolve as students gain clinical experience and as the GenAI tools themselves develop rapidly. The fast-paced evolution of the GenAI ecosystem means that the attitudes reported here represent a snapshot that may shift.

Second, the data rely on self-reporting, which is susceptible to social desirability bias. Furthermore, this study measures perceptions, reported willingness, and concerns, not objective clinical competency or the actual diagnostic precision of GenAI tools. There is a clear distinction between holding a positive attitude toward a technology and demonstrating proficiency in using it effectively and safely in practice.

Third, regarding the measurement instrument, the survey was designed as a descriptive item set for this specific exploratory purpose. While content validity was pursued through expert review and pilot testing, the tool has not undergone full psychometric validation (e.g., confirmatory factor analysis) as would be required for a formal scale. Its structural validity and generalizability require further testing.

Fourth, the use of convenience sampling, primarily through online and academic channels, limits generalizability. This method may have systematically over-represented students who are more technologically engaged or proactive in seeking digital resources, potentially inflating estimates of adoption rates and willingness. To mitigate this and enhance credibility within an exploratory framework, we sought a large and geographically diverse sample (n = 1062 from 168 institutions across 29 provinces). The demographic and academic heterogeneity of our sample (see Table 1) supports the internal diversity and national scope of this snapshot, though it does not constitute statistical representativeness. Consequently, and consistent with the study’s descriptive aim, we deliberately abstained from inferential statistical comparisons between subgroups, as such analyses would be inappropriate with a non-probability sample. Furthermore, some “Other” options could not be deeply analyzed due to the lack of specific information provided by the participants. Finally, it is important to note that our findings reflect attitudes within a specific national context at a given time. Variations in clinical learning environments and the level of perceived institutional support may also influence how students adopt and trust new educational technologies like GenAI [26, 27]. Therefore, the applicability of our conclusions to specific institutional settings should be considered with this contextual factor in mind.

Future research should build upon these exploratory findings. Longitudinal studies are needed to track attitude evolution. Experimental or quasi-experimental designs (e.g., introducing control groups) are required to assess the causal impact of GenAI on clinical reasoning. Probability sampling would enable valid subgroup comparisons and hypothesis testing. Multimethod approaches that combine surveys with behavioral observation (e.g., in simulated clinical scenarios) are crucial to bridge the gap between reported attitudes and actual competency. Finally, future work should develop and employ robust, validated scales to measure key constructs like AI acceptance, potentially based on the thematic dimensions identified here.

Abbreviations

AI

Artificial Intelligence

GenAI

Generative Artificial Intelligence

TAM

Technology Acceptance Model

Authors’ contributions

Study design: Xi Cao, Hao-Yue Gao. Literature screening and data collection: Xi Cao, Yu-Yao Lu, Jia-Hong Li, Xin-Yue Luo, Yu-Xin Zeng, Si-Heng Wang. Data analysis and manuscript preparation: Xi Cao, Hao-Yue Gao. All authors contributed to and have approved the final manuscript.

Funding

Financial support for this research were provided by Sichuan Research Center of Applied Psychology, Key Research Base of Philosophy and Social Sciences of Sichuan Province (foundation number: CSXL-22212); Chengdu Medical College-the Second Affiliated Hospital of Chengdu Medical College Joint Fund Project (foundation number: 2022LHFSSYB-06); Sichuan Medical Law Research Center 2022 Scientific Research Project (foundation number: YF23-Q15); Chengdu Medical College-the Second Affiliated Hospital of Chengdu Medical College Joint Fund Project (foundation number: 23LHHGYMP19). Research Center for Coordinating Urban and Rural Education Development, Key Research Base of Humanities and Social Sciences in Colleges and Universities of Sichuan Provincial Education Department (foundation number: TCCXJY-2023-B23). Key Natural Science Projects of Chengdu Medical College Science and Technology Fund for 2024 (CYZZD24-11).

Data availability

The datasets used and/or analyzed during the current study are available from the first author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Biomedical Ethics Committee of Chengdu Medical College, with the review opinion number: Chengdu Medical Ethics Review CMCIR2025NO.038. All participants obtained informed consent prior to inclusion. Our study adhered to the Declaration of Helsinki.

Consent for publication

Informed consent was obtained from all study participants. The questionnaire was anonymous, and the consent process included information about the use of their anonymized responses for research publication.

Competing interests

The authors declare no competing interests.