Physician Use of Large Language Models: A Quantitative Study Based on Large-Scale Query-Level Data

Introduction

Since its launch in late 2022, ChatGPT has been at the forefront of the burgeoning field of generative artificial intelligence (GenAI), sparking immense enthusiasm yet significant concerns within academic and professional communities [1-4].

GenAI models offer significant potential for innovation in the health care domain [5-11]. They have the potential to streamline time-consuming administrative tasks [12], support informed medical decisions through precise and personalized treatment recommendations [6], serve as an interactive encyclopedia for medical education [7], and drive medical research forward more generally [7]. However, the realization of these benefits depends on various factors. To provide value in clinical decision support, these technologies must first be integrated into existing technological systems [7] and their implementation must be coupled with significant modifications to medical workflows and processes [12]. Unfortunately, comprehensively implementing these technologies within the health care sector is made difficult because GenAI technologies present unique challenges [12]. The most notable risks that have been raised to date include data privacy concerns [12-14] (eg, noncompliance with the Health Insurance Portability and Accountability Act [15]), the concern that clinicians may enter patients’ protected health care information within queries, and “hallucination,” whereby LLMs could potentially mislead physicians [6].

As GenAI increasingly penetrates health care practice, it is essential to understand how physicians are adopting and using these technologies. Existing research on this topic generally falls into 3 main streams. The first stream is studying GenAI models for medical applications, as exemplified by various MedGPTs [16-19]. Although these studies demonstrate potential use cases for GenAI, they are typically conducted in experimental settings [16-19]. How physicians engage with these technologies in real-world settings remains poorly understood. The second stream involves small-scale qualitative research, such as interviews or surveys with limited participant pools [20,21]. Although valuable for exploratory insights, these studies offer limited generalizability due to their constrained sample sizes. The third stream consists of review papers that synthesize existing AI-related research to infer potential uses of GenAI in health care [22,23]. However, these reviews often lack detailed, practice-level insights into how GenAI is being integrated into clinical or research workflows. To date, few studies have quantitatively examined how physicians actually use GenAI in routine clinical or research practice.

Moreover, while many studies have highlighted the potential advantages of GenAI in health care [5-9], few address the challenges associated with its use. At this early stage of GenAI adoption, it remains unclear whether digital divides exist across different demographic groups. Digitally disadvantaged users may be less likely to adopt advanced GenAI tools, thereby hindering the widespread integration of these models. Additionally, given that GenAI systems are not yet formally regulated or integrated into mainstream electronic health record systems, privacy risks have not been systematically assessed.

Given the foregoing situation, this paper explored large-scale query-level data to quantitatively explore the use of GenAI by physicians in practice. Specifically, we sought to answer questions including the following: What types of topics do physicians inquire about? How do these topics evolve with increased engagement and vary across physician demographics? How frequently do the queries contain sensitive information?

Methods

Study Design

This study used a classification-based quantitative descriptive design, which is well-suited for examining use patterns and uncovering empirical insights into the use of GenAI [24,25]. We applied separate classification models to identify query topics and detect sensitive information—approaches that have demonstrated strong performance in processing large-scale textual data [26-28]. By categorizing the topics of physician queries, we were able to trace usage trends over time and across demographic groups. In parallel, our sensitivity analysis allowed us to evaluate potential privacy risks associated with physicians’ interactions with GenAI platforms.

Setting, Participants, and Sampling

Our query-and-answer data were collected from a company in Xiamen, China, which hosts a website that allowed physicians to access popular GenAI (eg, ChatGPT-4 and ChatGPT-3.5 from OpenAI) via these models’ application programming interfaces. This company invited physician participants from Dingxiangyuan, a leading medical community in China. As one of the largest web-based platforms for doctors, researchers, and health care professionals in the country, Dingxiangyuan provides a valuable resource for accessing a broad pool of physician users.

As most of these users were from China, they had limited access to cutting-edge GenAI models via non–application programming interface channels during our study period. The data collected are believed to reflect the majority of participating physicians’ use of GenAI during the observation period. Users’ demographic information, including age, gender, department, and device type, was also collected from this company.

Data Collection

This study collected 106,942 query-and-answer pairs by 989 physicians between August 29, 2023, and April 16, 2024, to examine how the use of GenAI evolved over time and across demographics. Approximately 40% (n=396) of the enrolled physicians were female, 45.9% (n=454) were younger than 25 years, and 54.1% (n=535) were between 25 and 56 years. The majority worked in clinical departments (n=680, 68.8%) or medical technology departments (n=127, 12.8%).

Data Analysis

We conducted two sets of text analyses. First, we classified the queries to identify the search topics from physician users. We integrated a GPT-4–based classifier with a rule-based method to perform this classification (for method details, see the Topic Classification section of Multimedia Appendix 1). These classified topics allowed us to determine the topics of physicians’ prevalent use cases and to examine if these use cases significantly varied with physicians’ engagement and their demographics.

Second, we performed an additional text-mining analysis to detect potentially sensitive information existing in the queries. We created a detector that incorporates regular expressions, Microsoft’s Presidio [29], and ChatGPT-based detection to thoroughly search for personally identifiable information (for method details, see the Sensitivity Detection section of Multimedia Appendix 1). By analyzing the detected sensitive information, we categorized and quantified the types and frequency of sensitive data present in physicians’ queries. This allowed us to assess the risks of privacy breaches and identify specific instances where breaches may occur.

Ethical Considerations

We obtained permission from the Xiamen Moniu Investment Company to use their dataset for our secondary data analysis. At the time of data collection, the company obtained informed consent from all participants, and all records were subsequently anonymized to safeguard privacy and confidentiality. Participants received complimentary access to the GenAI tool as compensation.

Results

Query Topics

To understand physicians’ objectives when they engage with GenAI models, we conducted text mining on a broad sample of input queries. Table 1 summarizes 15 topics extracted from user queries. They are divided into 4 broad groups in both health care–related and non–health care settings. As depicted, physicians used the GenAI tool not only for health care–related tasks, including clinical practice, medical research, and administrative work, but also for tasks unrelated to health care. The distribution across all 15 topic categories is also detailed in Table 1. In health care–related queries, a significant proportion of the 106,942 physician queries were related to medical research, such as writing and editing (n=24,688, 23.09%), translation (n=19,835, 18.55%), and data analysis and software use (n=17,412, 16.28%). This implies that physicians primarily use GenAI models for research in health care–related settings. Compared with conventional medical research tools like Google Scholar or PubMed, which serve primarily as search engines and databases for retrieving content, GenAI tools actively generate new text and help synthesize information across sources. Notably, physicians are not just consuming information. They are interactively collaborating with GenAI tools to think through research problems in real time (eg, following the chain of thought in solving data analysis problems). All of these advanced capabilities of GenAI are attracting users’ attention.

Further, our findings revealed a seeming reluctance among physicians to engage in queries related to clinical practice. Examples of queries and answers for clinical practice and medical research are presented in Textbox 1. These examples demonstrate that GenAI can provide reasonable answers to both types of questions. However, physicians searched for clinical practice significantly less frequently than they searched for medical research. Physicians sought information about basic medical science and clinical medicine at rates of 6.80% and 3.66%, respectively. Searches for medication information and surgical techniques were markedly lower, at just 0.94% and 0.85%, respectively. There are several potential underlying reasons for this hesitancy. First, physicians are cautious about sharing patient information on the third-party GenAI platform, even anonymously. Without sufficient patient information, queries related to clinical practice become significantly less effective. Second, the conversational nature of GenAI models requires substantial effort to sequentially input a complete patient health profile [30,31]. This is time-consuming and deters the integration of clinical practice with GenAI models. Third, the current capabilities of GenAI are limited in integrating complex medical information, often resulting in clinical ineffectiveness, particularly when multiple conditions must be considered [32].

In queries unrelated to health care, the majority tended to be informal. The largest proportion of these queries pertained to writing and editing tasks (12.86%), including the preparation of documents for non–health care purposes. An additional share of queries involved informal interactions (8.76%), such as searching for recipes, composing poetry, and writing greetings.

Usage Pattern

Our study also examined how physicians use GenAI models. For precision, we focused specifically on the 4 broad groups outlined in Table 1.

First, we analyzed the temporal dynamics of health care–related (including medical research, clinical practice, and administration groups) versus non–health care topics in user queries, as illustrated in Figure 1. The x-axis represents the percentile of physicians’ queries in sequential order. A higher x-axis value indicates a later query, while a lower value indicates an earlier one. The y-axis represents the proportion of queries related to either the health care or non–health care topic. This plot demonstrates an increase of around 20% in the proportion of health care–related queries and a corresponding decrease of around 20% in non–health care queries within approximately the first 15% of the query sequence, suggesting that engagement with the platform influences the predominance of health care queries. After this “warm-up” period, topic proportions stabilize, indicating a plateau in shifts as user engagement continues. This suggests that physicians start with casual, nonprofessional questions but shift to more serious ones over time, reflecting growing engagement and trust. These trends are also statistically verified through piecewise regression (see the Model Specification section of Multimedia Appendix 1).

‎

Second, we examined the heterogeneity in usage patterns across user demographics. Table 2 displays the group means across different dimensions. First, we observed that the usage patterns among physicians varied by gender. Compared with male physicians, female physicians asked a larger proportion of clinical questions (female: 0.154 vs male: 0.108; P<.001) and administration questions (female: 0.027 vs male: 0.018; P<.001). Second, age and experience also affected usage. Younger physicians, early in their careers, tended to pose more clinical-related questions (age ≤25: 0.146 vs age ∈ (25,40]: 0.115 vs age >40: 0.103; P<.001) compared to their senior counterparts, who were more inclined to inquire about research-related topics (age ≤25: 0.580 vs age ∈ (25,40]: 0.607 vs age >40: 0.664; P<.001). This suggests that, unlike younger physicians, experienced physicians seemingly rely more on their own expertise in clinical settings but seek external input more frequently when engaging in research activities. Third, usage patterns varied based on the device used. Physicians accessing the platforms via computers asked more research-oriented questions (computer: 0.637 vs mobile: 0.296; P<.001), whereas those using mobile devices primarily inquired about clinical (computer: 0.107 vs mobile: 0.264; P<.001) and non–health care (computer: 0.238 vs mobile: 0.393; P<.001) topics. This discrepancy indicates that clinical and casual queries often arise spontaneously and are more likely to be addressed on-the-go, while research activities are typically conducted in a more serious manner, for example, via using desktops. Fourth, the department (ie, specialty) of physicians mattered. Among all departments, staff from the clinical department asked the most research questions (clinical department: 0.634 vs medical technology department: 0.540 vs other department: 0.418; P<.001) and those from the medical technology department asked the most clinical questions (clinical department: 0.133 vs medical technology department: 0.141 vs other department: 0.078; P<.001). We further used regression analyses to examine the heterogeneity in usage patterns across user demographics. The results are shown in Table 3. We found that the heterogeneity in usage patterns remained consistent even after controlling for other demographic factors.

Data Privacy

We examined potential privacy issues arising from the use of GenAI models by physicians. As GenAI models typically work as third-party entities, they often caution users against inputting sensitive information into the system. However, a thorough analysis is essential to determine the extent to which physicians adhere to these guidelines and prevent inadvertent input of sensitive data.

According to the Health Insurance Portability and Accountability Act Privacy Rule, there are 18 types of identifiers that may contain sensitive or personally identifiable information [15]. As users were from China and had limited access to cutting-edge GenAI (Table 4), 15 types of identifiers were detected in our data.

We observed that users disclosed sensitive information on the GenAI platform extremely rarely. Among those rare instances, the most commonly disclosed sensitive information included human names, locations, and usernames, followed by ages, dates, and URLs. We further investigated the distribution of all types of identifiers across the 15 topic categories defined in Table 1. The majority of sensitive information was primarily derived from writing and editing. For instance, 576 of 958 instances of sensitive information on human names were linked to writing and editing, while 302 of 574 location-related sensitive information instances were connected to writing and editing. This suggests that users are less mindful of sensitive information when writing or editing documents. Moreover, sensitive information related to account names (435/460) and URLs (83/132) mainly arose from data analysis and software usage. This indicates a tendency for physicians to disclose sensitive information when copying directories or web sources.

To investigate how the privacy risk varied by user group, we regressed the indicator of whether to disclose sensitive information in the query on users’ demographic information. The results are shown in Table 5. In the first column, we used the ordinary least squares estimator, and in the second column, we used the logit estimator. Their results were consistent. We found that male physicians, compared with female physicians, were more likely to disclose sensitive information. Physicians from the medical technology department faced higher privacy risks than those from the clinical department and other departments. If physicians used the mobile version of the model, their likelihood of posting sensitive information was higher.

Discussion

Principal Findings

This paper leverages a comprehensive query-and-answer dataset to understand how physicians use the current third-party GenAI models. Our findings offer key insights into GenAI model usage and potential in health care. We identify that, when used by physicians, GenAI models are predominantly used for medical research tasks rather than direct clinical applications. This trend highlights a gap between the potential and actual uses of GenAI models in health care, particularly in clinical practice. Health institutions can begin by promoting GenAI use in research applications, such as documentation, literature review, or patient education. These uses can build trust among physicians and demonstrate early value. Once this foundation is in place, institutions should introduce formal validation frameworks, standardized performance benchmarks, and clinical guidelines to assess GenAI for higher-stakes, decision-critical use.

Another observation is the gradual increase in physicians’ reliance on GenAI, starting with informal or personal queries and then evolving to more professional queries. This progression indicates a growing trust and familiarity with the technology. Accordingly, with time and improved integration, GenAI models could become a more requisite tool in health care.

Further, the varied usage patterns based on demographic factors also reveal how different groups within the medical community are adapting to GenAI. These differences not only point to the existence of digital divides but also offer valuable insights into how user needs and behaviors vary by gender, age, device type, and departmental affiliation. First, we showed that female physicians ask a larger proportion of clinical questions and administration questions than male physicians. This suggests that female physicians are more likely to see GenAI as a practical assistant for managing clinical decision-making and routine administrative tasks. Second, we found that younger physicians depend on GenAI in clinical settings more than senior physicians do. Practically, younger physicians’ greater focus on clinical use may reflect higher openness to technology or a need for support with real-time decision-making, while senior physicians’ stronger interest in research may point to their roles in academic medicine. Third, our results showed that physicians using GenAI on computers tend to ask more research-related questions, while those using mobile devices focus more on clinical inquiries. Such device-based variation highlights the importance of context-aware interface design—mobile platforms may benefit from more streamlined, clinically oriented features. Last, we found that staff from clinical departments ask the most research questions, and those from medical technology departments ask the most clinical questions. This indicates that GenAI may be serving as a bridge between technical roles and clinical relevance, helping nonphysician staff understand how their work impacts patient care.

This study also quantifies privacy risks associated with GenAI, an area that has garnered increasing attention. Our findings indicate that while concerns about data privacy are acknowledged, the actual likelihood of sensitive information disclosure is relatively low. Importantly, we identify specific use scenarios—such as drafting or editing documents—where the risk of inadvertent disclosure is higher. These insights underscore the need for clear guidelines and safeguards to ensure responsible use of GenAI in medical contexts. Institutions should address this by establishing clear usage policies, input restrictions, and built-in safety mechanisms. For instance, institutions can enable input filters or redaction suggestions that flag or block sensitive content in real time, particularly for drafting or editing documents.

In sum, this real-world behavioral dataset offers a foundation for evidence-based governance of GenAI in health care. Regulatory agencies and professional bodies can use these insights to craft balanced guidelines that support innovation while safeguarding patient data and care quality.

Limitations

Although the results of this study are promising, there are a number of limitations that should be considered. First, the majority of participating physicians were based in China, which may limit the generalizability of the results to other health care systems. The health care infrastructure, digital health policies, and physician workflows in China differ significantly from those in other countries such as the United States or those in Europe. For example, in China, a notable proportion of GenAI use cases involves language translation—an application that may be less prevalent among physicians in English-speaking countries such as the United States. However, it is also important to highlight that many aspects of physician workflows are shared globally. Regardless of location, physicians commonly engage in both clinical decision-making and medical research, suggesting that core use cases for GenAI—such as writing and editing, basic medical science, and clinical medicine—are likely to be relevant across health care systems. Second, while this study uses a large-scale quantitative dataset, it does not incorporate qualitative data such as in-depth interviews or open-ended survey responses. Future studies could benefit from integrating qualitative methods to gain richer insights into physicians’ motivations, concerns, and contextual factors shaping GenAI usage. This mixed methods approach could also help explore nuances in trust-building and perceived risks that are difficult to capture through quantitative analysis alone. Last, the data used in this study were observational and derived from real-world usage logs, which means we cannot fully account for physicians’ intentions or the broader clinical context in which queries were made. This could limit the interpretability of some behavioral patterns.

Conclusions

This study reveals that physicians primarily use GenAI for research rather than clinical tasks, with usage patterns evolving alongside growing trust. Demographic differences shape adoption, and privacy risks remain low but nonnegligible. These findings provide insights into developing safe, effective integration of GenAI into health care practice.