This study used a cross-sectional survey design to explore clinicians’ knowledge and use of LLMs such as ChatGPT. The survey instrument was developed using a modified Delphi method by an expert panel of 10 physician informaticians from various pediatric specialties; survey questions were collaboratively codesigned with multiple rounds of feedback until consensus was achieved. The survey consisted of a series of structured and unstructured questions and used adaptive questioning; the full survey, including the adaptive questioning logic, is available in Multimedia Appendix 1. The survey was pilot-tested on a group of 5 pediatricians prior to being sent to the entire target population.

The study protocol, survey, and recruitment tool were granted full approval by the Boston Children’s Hospital Institutional Review Board (#P00045317). Participation in the survey was voluntary and uncompensated; respondents were informed of the purpose of the study and assured of their anonymity. The informed consent and privacy and confidentiality protection language is provided in Multimedia Appendices 1 and 2.

This closed survey study was conducted at Boston Children’s Hospital, a large academic urban pediatric health care system. The target sample was all Boston Children’s Hospital physicians and advanced practice providers in both hospital and outpatient clinic settings. Recruitment emails (see Multimedia Appendix 2) were sent via Research Electronic Data Capture (REDCap), a secure, web-based data capture application, which was also used for survey administration [18,19]. Survey recruitment and data collection started on October 12, 2023, and extended through November 14, 2023. Reminder emails were sent via REDCap to nonrespondents a maximum of 2 times.

Survey responses were analyzed in aggregate with minimum subgroup sizes of 10 responses to minimize the risk of reidentification of participants. Four survey questions included an “other” free-text option to capture concepts not covered by the provided answer choices. These free-text responses were analyzed qualitatively using the following methods. For each question, the predetermined discrete survey responses were used as a provisional codebook, which was expanded through inductive content analysis of the free-text responses [20]. To create the expanded codebook, 2 researchers (DJK and NR), one of whom was not involved with the data acquisition process, reviewed free-text responses and generated additional codes through an iterative process involving consensus meetings until no new codes were identified [21,22]. Coding conflicts were resolved by a third researcher (JDH). This expanded codebook was then applied to categorize free-text survey responses. Following coding, the 3 researchers (DJK, NR, and JDH) organized the coded responses into broader themes through an iterative process of consensus meetings.

The χ² test of independence, followed by a posthoc analysis of adjusted residuals, was used to assess differences in survey question responses across demographic variables, including age, gender, race/ethnicity, and clinical role.

Surveys were sent to a total of 3127 physicians and advanced practice providers via email; we received 390 (12.5%) completed survey responses. As shown in Table 1, most respondents self-identified as female (n=293, 76.3%), White or European (n=324, 83.7%), and either an attending physician (n=165, 42.4%) or advanced practice provider (n=110, 28.3%).

Among the 390 respondents, 288 (73.7%) indicated that they were familiar with ChatGPT or another LLM and an additional 83 (21.2%) indicated that they had heard of ChatGPT but did not really know what it is. Only 19 (4.9%) respondents reported that they had not heard of ChatGPT. Of those who had heard of ChatGPT (n=371), 197 (53.1%) had used ChatGPT or a similar model. Only 52 (26.7%) of 195 respondents who were using an LLM had used it clinically. Reported clinical uses of LLMs are shown in Table 2; the most common uses were drafting school and work letters and drafting prior authorizations.

Overall, 72.4% (142/196) of question respondents reported using LLMs for nonclinical work. Nonclinical use cases included drafting emails (55/139 39.6%); creating outlines for grants, papers, or teaching materials (52/139, 37.4%); drafting a letter of recommendation (51/139, 36.7%); and writing code (eg, for statistical analysis or data visualization) (18/139 12.9%). Respondent free-text uses that were listed in “other” are shown in Table 3.

Most respondents did not think ChatGPT should be used for patient care in its present state (256/390, 70.9%). Listed concerns were accuracy or reliability (319/390, 87.2%), patient privacy or security (237/390, 64.8%), unclear how ChatGPT makes decisions (232/390, 63.4%), lack of regulation (225/390, 61.5%), and potential bias in the data model (219/390, 59.8%). Free-text responses for this question are shown in Table 3.

Among the 367 respondents, 272 (74.1%) indicated they would use a version of ChatGPT if one were available that was compliant with the Health Insurance Portability and Accountability Act (HIPAA) [23]. Table 4 shows examples of how they envisioned using a HIPAA-compliant version of ChatGPT. If such a model were available, most participants indicated that they would feel comfortable entering patient diagnoses (223/266, 83.9%), age (188/266, 70.7%), and clinical questions without patient information (186/266, 69.9%); few would feel comfortable entering patient name (101/266, 38.0%), medical record number (87/266, 33%), date of birth (98/266, 37%), or whole notes from a patient chart (99/266, 37%). Table 3 shows the results of our analysis for free-text responses.

Trainees were more likely to have used ChatGPT than other respondents (P=.01); the percentage who endorsed using ChatGPT clinically was higher among trainees (40%) than nontrainees (24%), but the values in this question did not reach statistical significance (P=.06). Male respondents were also more likely to have used ChatGPT (P=.005). Respondents in the ≤29 years age group were more likely to be familiar with ChatGPT and those in the ≥70 years age group were less likely (P=.002). Responses to the rest of the survey questions did not significantly differ across demographics. Specifically, there was no statistically significant difference in whether an LLM was being used for clinical and/or nonclinical work, the endorsed current use cases, expressed concerns regarding LLM use in health care, desire for a HIPAA-compliant LLM, or endorsed planned uses for a HIPAA-compliant LLM.

This study demonstrates that pediatric providers are already using LLMs in both their clinical and nonclinical work. Because of the known limitations of LLMs in a clinical setting, including demonstrated low diagnostic accuracy, it is important to know how pediatricians are using these tools. This study adds to the current literature by providing granular information about how people are currently using LLMs in their work as well as detailing ways that providers envision using a HIPAA-compliant future version of ChatGPT.

Nearly all survey respondents had heard of ChatGPT, and most had used this tool. While nearly 75% of LLM users indicated that they are already using ChatGPT in their nonclinical work, only slightly over 25% indicated that they are currently using an LLM for clinical work. Moreover, the most common clinical uses reported were for administrative tasks such as drafting letters, prior authorizations, and patient education materials.

Similar to other early studies, we found that clinicians are enthusiastic about using LLMs (specifically, a HIPAA-compliant LLM) in clinical and nonclinical work. It is notable that almost one-third of pediatric providers in this study indicated that they would feel comfortable using LLMs for patient care in the current format. Additionally, almost three-quarters of respondents indicated that they would use a HIPAA-compliant version of ChatGPT if one were available. If a HIPAA-compliant LLM were available, participants described a variety of ways that they would use it. Most of the envisioned use cases for the HIPAA-compliant LLM were still administrative or operational; however, other common uses included drafting all or part of a clinical note as well as a variety of uses related to clinical decision support and clinical documentation.

Since we have demonstrated that LLMs are already being used clinically and that there is strong interest in further future use of LLMs, it is imperative that health care systems and government agencies create thoughtful policies and regulations for LLM use in health care. The Biden administration recently announced an executive order that directs actions to maximize the promise and manage the risks of AI [24]. Clinical informaticians are needed to help navigate the thoughtful implementation of AI tools into clinical care.

This study has limitations. Most notably, the generalizability is limited by the low response rate and pediatric provider population. Another limitation is selection bias, as providers using LLMs may be more interested in completing a survey related to this topic. Similarly, although recent American Board of Pediatric statistics show that 67% of pediatricians are female and 57% are White [25], our survey respondents were even more skewed toward these populations. Self-reported data may contain some social desirability bias as respondents may attempt to demonstrate that they are using these technologies in acceptable ways. Also of note, we did not ask if respondents worked in an inpatient, outpatient, or other setting; not having that information limits some interpretation of the response data. For example, the number of providers interested in using LLMs to write discharge summaries is less interpretable as these documents are not generally written by outpatient providers.

The enthusiasm we found for the future use of LLMs lends itself to further investigation of LLM use in health care. We propose evaluating the differences in use cases by clinical work setting such as the emergency department versus inpatient versus outpatient versus proceduralists. There would be value in determining if survey results would differ across practice settings such as nonacademic, nonurban, adult patients, and different patient resources, or by geographic location; thus, we propose a larger study across institutions in the future. In-depth interviews and other qualitative methods could be used to gain deeper insights into providers’ LLM use and beliefs. Finally, exploring patients’ perceptions and current use of LLMs would be of great value.

This survey study adds to the corpus of knowledge on how providers are thinking about and using LLMs in a clinical context. LLMs will add to the series of digital tools in the clinical ecosystem meant to help advance clinical care. Despite significant concerns and barriers to LLM use in health care, this survey demonstrates that these tools are already commonly used and there is enthusiasm for their future use. Knowing how providers are using LLMs in their clinical and nonclinical work will help guide policy and regulations regarding the health care use of AI. As informaticians, it is incumbent upon us to support the appropriate use of these technologies to improve patient care, while also monitoring for their unintended consequences.