Stratified sampling of participants
Survey was conducted during 16 May 2025–29 September 2025 in 14 provincial-level administrative divisions (“provinces”) of China. Per capita gross domestic product (GDP) in 2024 was > 15,000 US dollars ($) in 5 of the surveyed provinces (Jiangsu, Fujian, Zhejiang, Tianjin, Inner Mongolia), $10,001 –15,000 in 5 provinces (Anhui, Liaoning, Sichuan, Tibet, Shanxi), and $7001–10,000 in 4 (Qinghai, Henan, Hebei, Guangxi)11.
3493 physicians, including 2632 drawn from 32 tier-3 (highest-level) hospitals, 614 from 25 tier-2 hospitals, and 247 from 19 tier-1 (lowest-level) hospitals, participated in the survey (Fig. 1A; “Methods”). The surveyed hospitals are listed in Supplementary Table 1. 271 (7.8%) participants did not complete the survey.
Fig. 1: A stratified sample of 2708 physicians.
The alternative text for this image may have been generated using AI.
A Work flow for stratified sampling and data quality control. B Distribution of the 2708 respondents whose returned questionnaires passed data quality checks. “hosp.” and “phys.” stand for hospital(s) and physician(s), respectively.
Among the 3222 (92.2% of 3493) participants who completed the survey, median time to completion was 7 min (IQR, 4–12). To ensure data quality, we excluded data from 278 (8.6% of 3222) physicians who completed in <3 min. Subsequently, 236 (7.3% of 3222) filled-out questionnaires that failed to pass data quality checks were also excluded (“Methods”). The total “bad response rate” was 16.0% (514/3222; Fig. 1A).
The final dataset includes 2708 physicians (Fig. 1B). 1640 (60.6%) and 1068 (39.4%) are internists and surgeons, respectively. 980 (36.2%) are from the most economically developed provinces (Jiangsu, Fujian, Zhejiang, Tianjin, Inner Mongolia) in our sample, 954 (35.2%) are from moderately developed provinces (Anhui, Liaoning, Sichuan, Tibet, Shanxi), and 774 (28.6%) from the least developed provinces (Qinghai, Henan, Hebei, Guangxi). 2040 (75.3%), 490 (18.1%), and 178 (6.6%) respondents are from tier-3, -2, and -1 hospitals, respectively.
Demographics of respondents
The survey consisted of 32 questions (Q) divided into multiple sections (Supplementary Table 2; “Methods”).
Q1–Q6 collected the participants’ demographic information. 383 (14.1%) are chief physicians (highest-rank), with a male-to-female ratio of 1.3 (220/163), a median age of 51 years (IQR, 47–55), and median clinical practicing experience of 27 years (IQR, 22–31). 762 (28.1%) are associate chief physicians, with male-to-female ratio 1.2 (411/351), median age 44 (IQR, 40–48), and median experience 18 years (IQR, 14–24). 1563 (55.5%) are junior attending physicians (lowest-rank), with male-to-female ratio 0.8 (670/893), median age 36 (IQR, 33–39), and median experience 10 years (IQR, 7–14). Physician rank significantly correlates with sex (P < 0.0001 [chi-squared]), age (Spearman’s rank correlation ρ = 0.68 [P < 0.0001]), and years of clinical practice (ρ = 0.66 [P < 0.0001]).
Self-assessed levels of AI-related knowledge and use experience
Q7–Q10 asked the participants to self-assess their familiarity with AI and level of use experience. 1494 (55.2%) respondents indicated that they have some familiarity with AI-related knowledge and that they are confident to make decisions regarding AI technology based on their own judgment or after consulting with experts or perusing literature. 2311 (85.3%) respondents stated they have experience using medical AI. 1049 (38.7%) self-assessed they are good at using medical AI tools.
Male physicians are more likely to agree or strongly agree that they are good at using medical AI, with an odds ratio (OR) of 1.52 (95-percent confidence interval [CI], [1.29, 1.80]; P < 0.0001) compared with females after controlling for age, highest educational degree, years of clinical practice, weekly outpatient volume, weekly inpatient volume, physician rank, clinical specialty, hospital tier, and provincial per capita GDP.
Outlook for the future of AI in healthcare
Q11–Q15 asked the participants to express their outlook on the future of medical AI. The majority agree or strongly agree that AI will eventually transform the healthcare industry (n = 1951 [72.0%]), augment physicians’ capabilities (1934 [71.4%]), improve healthcare quality (1960 [72.4%]), improve equity in access to healthcare (1812 [66.9%]), or facilitate physicians’ training (1847 [68.2%]). 1954 (72.2%) respondents agree or strongly agree with ≥3 of these statements, and they have significantly higher self-assessed level of experience using medical AI (P < 0.0001) compared with the other respondents (n = 754 [27.8%]).
Q16 was a forced-choice question requesting the participants to consider the trade-off between model efficacy and physicians’ self-autonomy. 1782 (65.8%) respondents indicated they value their own autonomy more than an AI model’s clinical efficacy, implying that they prefer conditional to fully autonomous AI.
Projected use cases of AI-driven drug prescription
Q17 requested the participants to identify initial settings for drug-prescribing AI. 46 (1.7%) physicians responded “None”. The most commonly mentioned settings were: (1) when there exist standardized treatment guidelines, n = 1995 (73.7%); (2) when the decision is whether to extend a current therapy in someone, 1496 (55.2%); (3) when many variables need considering in the prescribing decision, 1198 (44.2%); and (4) when reliance solely on human decisions may cause critical delay in therapy, 660 (24.4%). Only 291 (10.7%) respondents indicated drug-prescribing AI is useful when there is a shortage of qualified staff. In other words, most physicians believe that, even with drug-prescribing AI, skilled physicians will remain indispensable.
Perceived importance of technical attributes
Q18–Q23 interrogated how physicians perceive the relative importance of 5 technical attributes of drug-prescribing AI: vetted efficacy, expediency, transparency, explainability, and governance/stewardship. Expediency and vetted efficacy received the highest average rank (1.9 [highest = 1, lowest = 5] and 2.0, respectively), significantly higher than explainability (3.5 [P < 0.0001]), governance/stewardship (3.7 [P < 0.0001]), and transparency (3.8 [P < 0.0001]).
For a drug-prescribing AI model to be perceived as expedient, 1645 (60.7%) respondents stated the model must be embedded in the clinical workflow, 1324 (48.9%) stated the model should shorten physicians’ work hours without compromising quality of care, 1099 (40.6%) stated the model should enable physicians to care for more patients, and 1056 (39.0%) stated the model should not increase physicians’ workload. Few respondents anticipate that drug-prescribing Al’s expediency lies in improving patients’ trust in physicians (n = 29 [1.1%]) or facilitating the hiring and training of junior physicians (102 [3.8%]).
To properly vet the efficacy of drug-prescribing AI, 1306 (48.2%) stated it is important that the model is developed using data from subjects similar to patients in their own practice, and 1068 (39.4%) stated the model needs to be validated in a domestic, Chinese cohort. In terms of model explainability, 1275 (47.1%) stated they need to understand how the model converts input variables to drug-prescribing recommendations, and 1273 (47.0%) respondents emphasize the model must explain its reasoning when its recommendation is discordant with physicians’ opinions. In terms of transparency, the majority of respondents (n = 2184 [80.6%]) indicated they need to know at least portions of technical details in model development before they can decide whether to adopt a drug-prescribing AI model in their clinical practice.
As for model governance and stewardship, 1468 (54.2%) indicated the importance of a trustworthy mechanism for monitoring quality, bias, and safety of AI prescriptions, 1244 (45.9%) emphasize mechanisms for model maintenance and/or upgrades, and 1167 (43.1%) desire easy accessibility to technical support staffs.
Perceived importance of institutional attributes
Q24–Q27 queried physicians’ perceived importance of hospital attributes in successful adoption of drug-prescribing AI. Most respondents indicated it is crucial that hospitals embrace new technologies (n = 1712 [63.2%]), value innovation (1711 [63.2%]), and invest resources in AI initiatives (1370 [50.6%]). In addition, 1358 (50.1%) stated that hospitals need to value employee satisfaction and workplace harmony, and 868 (32.1%) stated that hospitals need to promote learning and talent development.
For successful transition from the conventional drug prescribing workflow to AI-driven drug prescription, 1576 (58.2%) respondents highlighted the importance of having local “champions” for AI initiatives at hospitals, 1498 (55.3%) mentioned the importance of hospital leadership support, 1482 (54.7%) indicated that the hospital administration needs to provide an internal guidance on the use of drug-prescribing AI within the institution, and 887 (32.8%) stated that planning for the launch of drug-prescribing AI must be led by a multi-disciplinary task force.
The participants were also asked to give their opinions on what expertise(s) hospitals should focus on improving after adopting drug-prescribing AI. 1719 (63.5%) stated hospitals should improve ability to provide multi-disciplinary care to patients. 1278 (47.2%) and 1246 (46.0%) highlighted the importance of improving cost-effectiveness and throughput, respectively. 1093 (40.4%) stated hospitals should improve ability to manage difficult-to-treat diseases.
Perceived importance of governmental attributes
Q28–Q30 queried the participants’ opinions on the government’s role in successful adoption of drug-prescribing AI. Almost half (1317 [48.6%]) of the respondents prioritize a nation-wide policy mandate to promote AI in healthcare and a nation-wide initiative to build out technological infrastructures for medical AI over a nation-wide initiative to educate and train AI-related talents. 2037 (75.2%), 1861 (68.7%), and 1840 (67.9%) mentioned that professional medical societies, regulatory authorities, and separate individual hospitals, respectively, should be partially or wholly responsible for setting the standards for AI-driven drug prescription. 1472 (54.4%), 1221 (45.1%), and 1135 (41.9%) expressed that hospitals, AI developers, and physicians, respectively, should be remunerated for AI-driven drug prescription.
Self-assessed readiness for adopting drug-prescribing AI
Q31 and Q32 asked the participants to assess their hospitals’ and their own readiness for adopting drug-prescribing AI. 2224 (82.1%) forecasted their hospitals would be ready to adopt drug-prescribing AI within 5 years. 2117 (78.2%) forecasted they themselves would be ready to adopt within 5 years.
Clustering analysis of psychological profile-types
Clustering analysis of the respondents’ psychological characteristics (i.e., responses to Q11 – Q32) indicated that the surveyed physicians can be classified into 2 distinct psychological profile-types (Fig. 2). We termed the 2 psychological profile-types “optimists” (n = 1358 [50.1%]) and “pragmatists” (1350 [49.9%]). To assess the robustness of the clustering result, we performed 1000 independent bootstrapping experiments. Optimists’ average proportion in the bootstrapping experiments was 64.5% (standard deviation [SD] = 16.2%), and pragmatists’ average proportion was 35.5% (SD = 16.2%).
Fig. 2: Identification of Chinese physicians’ psychological profile-types with respect to drug-prescribing artificial intelligence (AI).
The alternative text for this image may have been generated using AI.
The heat map on the left-hand side, wherein each row represents a separate survey item and each column a separate respondent, is the result of hierarchical clustering on the 2708 respondents based on their psychological characteristics (i.e., responses to Q11 – Q32 [yellow, lower values; red, higher values]; Methods). Survey questions Q11 – Q32 are listed on the right-hand side.
How physicians of different psychological profile-types responded to survey questions is summarized in Supplementary Table 2.
There is a positive but insignificant correlation between provincial per capita GDP and the proportion of optimists among the physicians in a province: Spearman’s rank correlation ρ = 0.16 (P = 0.59). Proportions of optimists are comparable among physicians working at tier-3, -2, and -1 hospitals (51.1% [1042/2040], 46.3% [227/490], and 50.0% [89/178]; P = 0.17; chi-squared test). Proportions of optimists in surgeons and internists are comparable (51.6% [551/1068] vs. 49.2% [807/1640]; P = 0.24).
Most physicians, regardless of psychological profile-type, are receptive to using autonomous drug-prescribing AI in the near future. However, a higher proportion of optimists, compared with pragmatists, forecasted they will be ready to adopt drug-prescribing AI within 5 years (81.5% [1107/1358] vs. 74.8% [1010/1350]; P < 0.0001).
A higher proportion of optimists agree or strongly agree that AI will eventually transform the healthcare industry (79.7% [1082/1358] vs. 64.4% [869/1350]; P < 0.0001 [chi-squared]), augment physicians’ capabilities (79.8% [1084/1358] vs. 63.0% [850/1350]; P < 0.0001), improve healthcare quality (80.9% [1098/1358] vs. 63.9% [862/1350]; P < 0.0001), improve equity in access to healthcare (77.0% [1046/1358] vs. 56.7% [766/1350]; P < 0.0001), or facilitate physicians’ training (78.8% [1070/1358] vs. 57.6% [777/1350]; P < 0.0001) compared with the pragmatists. Single-variable logistic regression indicated that optimists are significantly more likely to value model efficacy above physicians’ self-autonomy (OR = 1.65; 95-percent Cl, [1.40, 1.94]; P < 0.0001) compared with the pragmatists.
Optimists and pragmatists have different priorities and expectations for drug-prescribing AI models. A significantly higher proportion of optimists, compared with pragmatists, stated that drug-prescribing AI would be most useful when there are standard treatment plans according to clinical guidelines (80.8% [1097/1358] vs. 66.5% [898/1350]; P < 0.0001). Optimists are also more tolerant of “black-box” AI models and therefore less likely than pragmatists to strongly insist AI explain the rationale underlying the prescription it recommends (12.6% [171/1358] vs. 21.5% [290/1350]; P < 0.0001). When considering the expediency of a drug-prescribing AI model, the majority (74.6% [1013/1358]) of optimists insist the model be embedded in the clinical workflow, whilst most (58.9% [795/1350]) pragmatists stated the model should shorten physicians’ work hours. With respect to vetting a drug-prescribing AI model’s efficacy, optimists focus more on the characteristics of patient cohorts used to develop and test the model. Optimists, compared with pragmatists, rely less on endorsements from high-impact-factor academic journals (20.0% [272/1358] vs. 33.3% [450/1350]; P < 0.0001), international medical professional societies (15.1% [205/1358] vs. 24.8% [335/1350]; P < 0.0001), or domestic medical societies (11.2% [152/1358] vs. 20.1% [271/1350]; P < 0.0001). In terms of model explainability, most (57.1% [775/1358]) optimists indicated that it is important they understand how input variables are converted to model outputs, whilst the majority (64.9% [876/1350]) of pragmatists insist the model explain its reasoning when giving a recommendation discordant with their opinions. Regarding model governance and stewardship, optimists emphasize a trustworthy mechanism for model maintenance and upgrades (60.2% [817/1358]) whilst pragmatists, a mechanism for monitoring quality, bias, and safety of AI prescriptions (68.4% [924/1350]).
The 2 psychological profile-types also have different expectations for the role played by hospitals in successful adoption of drug-prescribing AI. Regarding institutional culture, the majority of optimists emphasize that hospitals need to value innovation (80.3% [1090/1358]) and embrace new technology (76.0% [1032/1358]), whereas a higher proportion of (45.2% [610/1350]) pragmatists, compared with optimists (19.0% [258/1358]), emphasize that hospitals need to promote learning and talent development. To successfully implement drug-prescribing AI, the majority of optimists emphasize leadership of the hospital administration (72.2% [981/1358]), whilst pragmatists most emphasize the roles played by local AI “champions” (51.9% [701/1350]) and a multi-disciplinary task force (46.0% [621/1350]).
Optimists and pragmatists have different outlooks on their clinical practice after adopting drug-prescribing AI. Optimists, compared with pragmatists, are more likely to view drug-prescribing AI as an opportunity to enable them to focus on improving patient care throughput (66.9% [909/1358] vs. 25.0% [337/1350]; P < 0.0001), cost-effectiveness (63.0% [856/1358] vs. 31.3% [422/1350]; P < 0.0001), and empathetic caregiving (44.2% [600/1358] vs. 17.2% [232/1350]; P < 0.0001). In contrast, pragmatists stated they would emphasize the improvement of multi-disciplinary care (76.5% [1033/1350]), management of difficult-to-treat diseases (54.1% [730/1350]), and therapeutics research (44.5% [601/1350]) after adopting drug-prescribing AI.
Finally, the 2 psychological profile-types have different expectations for the government’s role in successful adoption of drug-prescribing AI. Optimists, compared with pragmatists, more emphasize the importance of a nation-wide policy mandate to promote AI in healthcare (P < 0.0001), whereas pragmatists, compared with optimists, more emphasize a nation-wide initiative for cultivating talents for medical AI (P < 0.0001). A higher proportion of optimists, compared with pragmatists, believe that professional medical societies and/or hospitals alone (that is, without involvement of regulatory authorities) could set the standards for drug-prescribing AI (38.9% [528/1358] vs. 23.6% [319/1350]; P < 0.0001). In contrast, the majority (54.2% [732/1350]) of the pragmatists indicated that standards should be set jointly by regulatory authorities, professional medical societies, and hospitals.
Mechanisms underlying the genesis of psychological profile-types
Subsequently, we investigated potential mechanisms underlying the genesis of the 2 psychological profile-types.
Multiple-variable logistic regression considering demographic co-variates (i.e., responses to Q1–Q6) and environmental factors, including hospital tier and provincial per capita GDP indicated that male physicians (OR = 1.46; 95-percent CI, [1.24, 1.72]; P < 0.0001) are more likely to be optimists compared with female physicians. The other co-variates including age, highest educational degree, clinical specialty, clinical practicing experience, physician rank, and provincial per capita GDP do not significantly contribute to the determination of psychological profile-type (Fig. 3A).
Fig. 3: Determinants of Chinese physicians’ psychological profile-types with respect to drug-prescribing artificial intelligence (AI).
The alternative text for this image may have been generated using AI.
A Effects of demographic co-variates and environmental factors (hospital tier and provincial per capita GDP) on psychological profile-type. Estimates and 95-percent confidence intervals (CIs) of odds ratios (ORs), calculated without bootstrapping, are displayed. GDP stands for gross domestic product and is expressed in US dollars ($). B The male sex’s effect on the probability of being an optimist mediated by medical AI use experience. P(Optimist) stands for the probability of being an optimist. ACME and ADE stand for average causal mediation effect and average direct effect, respectively. Displayed estimates and 95-percent CIs of ACME, ADE, and the proportion of mediation were calculated without bootstrapping. ACME and ADE passed significance (P < 0.05) in 98.3% (Supplementary Fig. 1) and 66.7% of 1000 independent bootstrapping experiments, respectively. C Independent effects of demographic co-variates, environmental factors, and self-assessed level of medical AI use experience on psychological profile-type. Estimates and 95-percent CIs of ORs, calculated without bootstrapping, are displayed. Independent effects of sex and level of medial AI use experiences passed significance in 66.5% and 98.9% (Supplementary Fig. 1) of 1000 bootstrapping experiments, respectively.
Causal mediation analysis indicated that the male sex’s total effect on the probability of being an optimist is 9.2 percentage points (95-percent CI, [4.9, 13.3]; P < 0.0001) compared with females, and 27.2% (95-percent CI, [12.8%, 49.3%]) of this total effect is mediated by the physician’s self-assessed level of experience using medical AI. Put otherwise, the average causal mediation effect (ACME) is 2.51 percentage points (95-percent CI, [1.11, 3.75]; P < 0.0001) increase in the probability of being an optimist (Fig. 3B). In contrast, physician rank is not a significant mediator of the male sex’s effect on the probability of being an optimist (ACME = –0.04 percentage points; 95-percent CI, [–0.30, 0.20]; P = 0.72) even though physician rank correlates with sex.
Finally, including medical AI use experience in multiple-variable logistic regression, we found that a high self-assessed level of using medical AI is associated with an OR of 2.98 (95-percent CI, [2.53, 3.51]; P < 0.0001) for being an optimist compared with physicians with a low level of use experience (Fig. 3C).
To assess the robustness of our findings, we ran 1000 independent bootstrapping experiments, each time re-sampling the 2708 respondents with replacement and repeating the clustering analysis, causal mediation analysis, and multiple-variable logistic regression. Our main results remained qualitatively the same (Supplementary Fig. 1). The male sex’s ACME via medical AI use experience passed significance (P < 0.05) in 98.3% of the bootstrapping experiments. In other words, medical AI use experience is a robust, significant mediator of sex’s effect on psychological profile-type. In contrast, sex’s average direct effect on psychological profile-type did not pass significance in 33.3% of the bootstrapping experiments. The independent effect of medical AI use experience on psychological profile-type is robust, passing significance in multiple-variable logistic regression in 98.9% of the bootstrapping experiments; the average OR for being optimists was 2.55 (SD = 0.66) in physicians with high vs. low self-assessed levels of medical use experience, after controlling for demographic co-variates and environmental factors.