Attitudes Toward Large Language Models in Health Care and Preferences for Their Adoption and Oversight Among Health Care Professionals: Cross-Sectional Survey
Synopsis
This cross-sectional survey, distributed online through a health care news platform mailing list, gathered responses from 335 health care professionals (including 230 attending physicians, 68.7%) and found that 62.7% reported current or contemplated large language model use, users reported significantly higher self-reported knowledge than nonusers (P < .001), the most valued applications were literature review (73.4%), decision support (57%), and patient communication (54.9%), leading concerns were decision errors (75.5%) and algorithmic bias (73.1%), 96.4% expressed concern about bias with those who had observed bias reporting higher concern (P < .001), and respondents favored regulation by professional associations (65.4%) over technology companies (29%), with 87.
Large language model (LLM) use rates by health care professional role among 335 survey respondents. Bars represent the percentage of respondents within each role who reported using (blue) or not using (red) LLMs. Other roles include medical physicist, clinical psychologist, doctor of pharmacology, health care executive, registered nurse, researcher, physical therapist, and medical scribe.
PubMedInterpretation
Reports early adoption patterns of large language models among health care professionals: 62.7% (n=210) reported current or contemplated use, users reported significantly higher self-reported knowledge than nonusers (P < .001), and age was not associated with knowledge (ρ=-0.072; P=.19). Prior empirical data on health care professionals' perceptions were limited; this survey of 335 professionals characterizes the association between use and knowledge and indicates that age is not a source of knowledge differences. Cross-sectional convenience sample of 335 respondents, analyzed with Wilcoxon rank-sum tests and Spearman correlations, reporting P values and a correlation coefficient.
Identifies the applications respondents considered most valuable as literature review (n=246, 73.4%), decision support (n=191, 57%), and patient communication (n=184, 54.9%), while leading concerns centered on decision errors (n=253, 75.5%) and algorithmic bias (n=245, 73.1%). Presents perceived value and safety concerns side by side, indicating that professionals lean toward adoption for lower-risk tasks while remaining cautious about higher-risk steps. Descriptive statistics reporting counts and percentages of endorsement for each option.
Finds that bias concern was near-universal: 96.4% (n=323) expressed concern about bias, and those who had observed bias reported higher levels of concern (P < .001). Links bias concern to personal observation experience, suggesting a statistical association between direct exposure to bias and the degree of concern. Group comparison reporting P < .001 within a sample of 335 respondents.
Reveals oversight preferences: 65.4% (n=219) favored regulation by professional associations versus 29% (n=97) favoring technology companies, 87.8% (n=294) supported professional guidelines, and 66.6% (n=223) reported no confidence in existing oversight. Provides evidence on how professionals rank governance bodies, showing a preference for professional organizations over industry and identifying a confidence gap in current oversight. Descriptive statistics reporting counts and percentages of preference distribution.
Perspective
This survey is suited to describing the attitudes of this particular convenience sample of health care professionals, composed mainly of attending physicians aged 30 to 59 in the Northeast United States and recruited through a health care innovation-focused news platform mailing list. The authors note that given the low response rate and recruitment channel, these findings may not reflect the views of the broader health care professional population, so the results are better treated as a starting point for implementation planning and further research than as an inference about all professionals.
Readers may still watch: how the low response rate and convenience sampling affect representativeness; the gap between self-reported knowledge and actual capability; the unclear direction of the association between use and knowledge; whether the link between bias concern and observation experience is influenced by other factors; and whether future studies with more representative sampling reproduce these preference distributions.