Health equity and public acceptance of large language models in healthcare in China: A national population-based survey

Large language models (LLMs) are entering healthcare, but their ability to improve access depends on public willingness to use them. If acceptance is socially patterned, deployment may widen existing inequities. We quantified public acceptance of LLMs in healthcare and examined its bio-psycho-social correlates in China. We conducted a representative, multistage stratified survey of adults aged 18 years or older across 150 Chinese cities between June and September 2024. After a standardized description of LLM capabilities, participants rated acceptance of LLMs in healthcare on a 0–100 scale. Survey-weighted regression identified correlates of acceptance, and Classification and Regression Tree (CART) analysis identified profiles associated with non-acceptance. Among 35,861 respondents, weighted mean acceptance was 64.3/100 (95% CI 63.9–64.6). Acceptance was lower among respondents with chronic conditions than in the overall sample (62.4 vs 64.3) and declined with age. Higher acceptance was associated with greater perceived social status, prior digital health use, self-efficacy, stronger family-neighbour relationships, and higher eHealth literacy (β_std = 0.06 to 0.19), whereas lower acceptance was associated with older age, adverse social and developmental exposures, severe ADHD symptoms, social loneliness, and financial strain (β_std = -0.04 to -0.09). In CART analysis, lack of prior digital health use defined the largest non-acceptor group, comprising 62% of the sample. Among those with prior digital experience, lower social support and lower childhood socioeconomic status remained important markers of non-acceptance. In the test set, the tree showed modest discrimination (weighted AUC 0.62, 95% CI 0.61-0.64), with high specificity (0.94) and low sensitivity (0.19). Public acceptance of LLMs in healthcare in China was moderate but unequal. Lower acceptance among older adults, people with chronic conditions, and those with fewer social and digital resources suggests socially patterned uptake. Equity-oriented implementation strategies are needed so LLM integration does not preferentially benefit advantaged groups.Author summary: Large language models are rapidly entering healthcare, but their value will depend on whether the public is willing to use them. We examined public acceptance of these tools in China in a national survey of 35,861 adults. The mean acceptance score was 64.3 out of 100, suggesting moderate but far from universal acceptance. Overall, acceptance was moderate rather than high. We found that willingness to use large language models in healthcare was not evenly distributed across the population. Older adults, people with chronic conditions, and those with fewer social and digital resources were less accepting. In contrast, people with stronger confidence, better digital experience, and greater social support tended to be more accepting. We also found that lack of prior digital health use was the strongest marker of non-acceptance. These findings matter because new health technologies do not automatically reduce inequality; they can also deepen it if they mainly benefit people who are already better resourced. Our study shows that the future implementation of large language models in healthcare should pay close attention to fairness, accessibility, and support for groups that may otherwise be left behind.