Abstract

Introduction:

China’s vigorous advancement of the “Internet Plus Healthcare Services” initiative, as a core component of social security system optimization for improving residents’ health welfare, furnishes a robust impetus for the integration of Generative Artificial Intelligence (GenAI) within the healthcare sector, particularly in the domain of health information provision, where GenAI is widely acknowledged to harbor substantial transformative potential. Nevertheless, empirical research specifically probing the modalities through which users adopt GenAI for health information seeking remains limited in extant scholarly literature.

Methods:

This study garnered primary data via a structured online survey and employed partial least squares structural equation modeling (PLS-SEM) to dissect the antecedent factors and underlying mechanisms governing users’ health information seeking intention via GenAI, grounded in the perspective of user perceptions.

Results:

PLS-SEM results indicate that user perceptions of GenAI (perceived competence, perceived convenience, and perceived anthropomorphism) positively influence on both user trust in GenAI and subjective norms, which in turn positively affect users ‘health information seeking intention through GenAI. Moreover, digital health literacy significantly moderates the relationship between user perceptions of GenAI and their health information seeking intention.

Discussion:

These findings yield valuable empirical insights for facilitating the optimized and scaled adoption of GenAI in health information services, enhancing the public health output of digital health policies, improving residents’ health welfare, and further alleviating the operational burdens borne by traditional healthcare resources.

1 Introduction

In recent years, China has successively introduced a series of policies, including Healthy China 2030 Plan1 and the Internet Plus Healthcare initiative,2 and identified digital health services as a key measure to improve the social security system and enhance residents’ health welfare. Health information refers to all types of information related to physical and mental health, disease prevention, nutrition, and wellness maintenance. Traditionally, the public has primarily sought health information from medical professionals (1). However, the rapid advancement and widespread adoption of digital technologies have fostered a new paradigm of health information-seeking that integrates intelligent technologies intrinsically. Among them, generative artificial intelligence (GenAI) represents one of the most promising modes for acquiring health information (2, 3). GenAI refers to a category of artificial intelligence models capable of generating new content based on patterns learned from large-scale training data. Unlike traditional rule-based or discriminative AI systems, GenAI leverages advanced architectures such as large language models to produce human-like and contextually relevant content. In the medical and healthcare field, GenAI has rapidly evolved from a niche technological innovation into a widely accessible tool for providing health information. This trend is mainly driven by the maturity of publicly available models such as ChatGPT and DeepSeek, as well as the growing public demand for instant, low-cost health advice. Endowed with distinctive advantages—including providing real-time responses to intricate queries (4), generating personalized content tailored to individual needs (5), and offering cost-effective, round-the-clock health knowledge services (6), GenAI has transcended the inherent limitations of traditional health information-seeking approaches. By enabling the public to obtain reliable health information without direct medical consultation, GenAI can effectively divert non-urgent health inquiries away from grassroots healthcare facilities, thereby reducing their consultation pressure. This function directly alleviates the operational burden on traditional medical resources and aligns with the core objective of the Internet Plus Healthcare initiative to optimize the social security system.

Unlike conventional health information platforms, GenAI is capable of deeply deciphering users’ latent needs and delivering customized, structured, and readily comprehensible health advice. This salient feature substantially lowers the threshold for the public to access professional health information. Accurate and accessible health information can encourage the public to adopt evidence-based and appropriate health behaviors, thereby providing a robust impetus to China’s vigorously promoted Internet Plus Healthcare initiative and the national Healthy China strategy. The implementation of these policies and empowerment of technologies have not only promoted the digital transformation of health information services, but also directly improved health accessibility for residents—especially vulnerable groups. These efforts have exerted a profound impact on enhancing the public’s mental health, life satisfaction, and sense of happiness, providing a new practical path for China’s social security system to improve residents’ health welfare. The global AI-driven health information service market is currently undergoing a period of explosive growth. Industry projections indicate that the scale of the global artificial intelligence market will surpass $1.3 trillion by 2032, underscoring the immense economic value embedded in this domain (7). During the COVID-19 pandemic, the suspension of offline medical services and prevalent public concerns regarding virus transmission expedited the migration of health information-seeking channels from traditional offline venues to online platforms. This paradigm shift in consumption and information-seeking behaviors has endured into the post-pandemic era, further catalyzing the application and popularization of GenAI in the health information sector.

However, for GenAI users, the rapid market expansion is accompanied by a series of challenges, including concerns over accuracy, requisite digital skills, and access barriers (8). The internal operational mechanisms of GenAI remain opaque, and the accuracy of its outputs cannot be guaranteed (9). In the context of health information provision, the proliferation of inaccurate and high-risk content and the hallucination risks inherent to GenAI have negatively impacted users’ intention to adopt this technology (10–12). Public trust in GenAI—particularly regarding complex health-related issues—remains relatively low (13). In addition, potential risks such as privacy breaches, algorithmic bias, and inconsistent professionalism in intelligent responses have aroused widespread concern among users (14, 15). For individual users, these multifaceted factors collectively influence their intention to use GenAI for health information seeking, leading to significant heterogeneity in usage behavior. Empirical evidence indicates that medical experts and online search engines remain the most prevalent sources of health information, while the adoption of GenAI in this domain remains limited (16). Therefore, exploring the key antecedents and internal mechanisms that drive users to adopt GenAI for health information seeking is of great significance for promoting the sound development of the GenAI health service industry, meeting the public’s diversified health information needs, helping China’s social security system improve residents’ health welfare through digital health services, and deepening the public health outcomes of the Internet Plus Health Services program.

Existing research on users’ health information-seeking intention or behavior has predominantly explored its influencing factors from the lenses of users’ individual attributes, online platform features, and broader contextual factors. Users’ internal perceptual constructs (e.g., performance expectancy, perceived susceptibility, effort expectancy, perceived benefits, perceived severity, perceived barriers, and self-efficacy), health literacy, habits, and hedonic motivation all exert significant impacts on users’ adoption intention toward health information-seeking tools (16–19). On the platform dimension, relevant studies have focused primarily on the inherent attributes of platform information and the features of access channels—such as information practicality, perceived credibility, platform reliance willingness, and technical specifications—which jointly determine individuals’ propensity to select a specific channel for health information acquisition (20, 21). Contextual factors that modulate individuals’ health information-seeking intention encompass disease stigma, norms governing information-seeking behaviors, social influence, and cultural underpinnings (22, 23). Notably, a subset of studies has posited that health literacy and social influence exhibit no significant correlation with users’ behavioral intention to adopt generative AI for health information-seeking purposes (16).

However, extant studies have largely focused on users’ health information-seeking behaviors via conventional online tools, while overlooking the impacts of GenAI’s unique perceived attributes on users’ health information-seeking intention. Prior applications of the unified theory of acceptance and use of technology (UTAUT) and technology acceptance model (TAM) in health information contexts have primarily examined generic constructs such as perceived usefulness and ease of use (16, 17). Yet they have largely neglected GenAI-specific attributes such as anthropomorphism and instantaneous responsiveness that differentiate GenAI from conventional search engines and health apps (6). With its distinctive advantages in interactivity, intelligence, personalization, and convenience, GenAI renders users’ perceived experience a core antecedent of their usage intention. In the context of GenAI-powered health information services, users’ active perceptions of the technology’s core attributes directly shape their trust in the disseminated information and subsequent usage behaviors. Therefore, it is imperative to integrate the user perception perspective into the research framework of GenAI-mediated health information-seeking behavior, address the gaps in the existing literature, and deepen the understanding of the mechanisms underpinning users’ adoption of intelligent technologies for health information seeking.

To address this research gap, the present study focuses on the impacts of users’ perceptions of GenAI’s core technical attributes on their health information-seeking intention. Specifically, this study examines the effects of three key perceived attributes—perceived competence (PCP), perceived convenience (PCV), and perceived anthropomorphism (PAN)—on users’ willingness to seek health information via GenAI. Furthermore, we test the mediating roles of trust in GenAI and subjective norms (SN) in the aforementioned relationships. Meanwhile, this study explores the moderating effect of digital health literacy (DHL) on the link between users’ perceptions of GenAI and their GenAI-enabled health information-seeking intention (HISI).

This study makes three key contributions to the extant literature. First, it extends the research agenda on how GenAI’s perceived technical attributes influence users’ health information-seeking intention. Existing studies rarely systematically delineate the core perceived technical dimensions of GenAI, nor do they investigate the differential effects of these dimensions on users’ information-seeking intention in the highly sensitive and professional context of health services. By focusing on three core perceived technical attributes among Chinese users and their impacts on health information-seeking intention, this study enriches the scholarly discourse on user technology acceptance and health information behavior within the domain of digital health communication. Second, this study further unpacks the specific mechanisms through which GenAI’s perceived technical attributes affect users’ health information-seeking behaviors, aiming to address the limitation that existing studies have overlooked the potential mediating pathways between technology perception and health information-seeking intention. By systematically examining the mediating roles of trust in GenAI and subjective norms, this study refines the theoretical model elucidating the driving mechanisms of users’ behavioral intention in GenAI-enabled health information contexts. Finally, this study investigates the moderating effect of digital health literacy on the relationship between users’ perceptions of GenAI and their health information-seeking intention. This inquiry may provide a novel perspective for reconciling the inconsistent findings reported in the literature regarding the link between technology perception and users’ information-seeking behavioral intention.

2 Hypothesis development2.1 User perception and health information seeking intention

Drawing on the Unified Theory of Acceptance and Use of Technology (UTAUT), the perceived usefulness and ease of use of emerging technologies exert a notable positive impact on the frequency of technology adoption, which in turn fosters the development of users’ behavioral intentions (24). The rationale for GenAI being regarded as a “potential ally” in the healthcare service domain stems from its acknowledged functional capabilities and the perceived accuracy of the health information it disseminates (25). Perceived competence is defined as users’ subjective perception of the professionalism and efficiency demonstrated by GenAI in executing health-related tasks (26). This construct encapsulates dimensions of intelligence, competence, technical proficiency, and operational efficiency (27). The stronger users’ perceived competence of GenAI, the more pronounced their intention to utilize this technology for health information-seeking purposes. Perceived convenience is recognized as a pivotal driver of users’ GenAI adoption, as well as a salient factor influencing their health information-seeking intention (17, 28). Perceived anthropomorphism reflects GenAI’s human-like interactional features, which can readily elicit users’ emotional resonance, mitigate their sense of unfamiliarity with intelligent systems, and enhance the overall user experience of interaction. Notably, familiarity and interactive enjoyment constitute core user-valued attributes that shape the adoption intention of digital health tools (29, 30). Extant empirical studies further indicate that the anthropomorphic traits of artificial intelligence can exert a positive influence on users’ adoption intention and facilitate the uptake of GenAI applications in practical scenarios (7, 31). Based on the above analysis, the following hypotheses are proposed:

H1: Perceived competence positively affects HISI.

H2: Perceived convenience positively affects HISI.

H3: Perceived anthropomorphism positively affects HISI.

2.2 The mediating role of trust

Trust is conceptualized as users’ attitudinal orientation toward GenAI. Trust in GenAI is manifested in positive expectations and evaluations of the technology’s performance, operational processes, and intended purposes, as well as the willingness to accept potential vulnerabilities and assume risks based on health recommendations provided by GenAI (5, 32). When users hold stronger perceptions of GenAI’s perceived competence, perceived convenience, and perceived anthropomorphism, they are more inclined to recognize the accuracy and professionalism of the health information generated by the technology, thereby fostering a sense of reliance and trust in GenAI. Extant studies have demonstrated that the anthropomorphic traits of artificial intelligence can exert a salient positive effect on user trust (33). In an empirical investigation of the antecedents of consumer trust in AI chatbots, Li et al. (34) further verified that anthropomorphism exerts a significant impact on user trust. Trust constitutes a critical antecedent of users’ willingness to adopt AI-enabled tools; specifically, users with higher levels of trust in GenAI are more likely to use it for health information seeking (35, 36). Therefore, we propose:

H4a: Perceived competence positively influences HISI through trust.

H4b: Perceived convenience positively influences HISI through trust.

H4c: Perceived anthropomorphism positively influences HISI through trust.

2.3 The mediating role of subjective norms

Subjective norms are defined as the extent to which individuals perceive that significant others believe they ought to adopt a specific system or technology (37). In the early adoption phase of intelligent tools, external validation tends to take precedence over personal experiential knowledge (38). When users hold favorable perceptions of GenAI’s competence, convenience, and anthropomorphism, they are more susceptible to the endorsement and recommendations of family members, friends, and other significant stakeholders. Consequently, such users perceive greater external normative pressure to utilize GenAI for health information-seeking purposes. In other words, users’ positive perceptions of GenAI exert a positive influence on their subjective norms regarding GenAI-based health information seeking. Furthermore, subjective norms have been empirically verified to exert a direct impact on users’ behavioral intention to adopt GenAI-powered chatbots (39). Based on the above, we propose:

H5a: Perceived competence positively influences HISI through subjective norms.

H5b: Perceived convenience positively influences HISI through subjective norms.

H5c: Perceived anthropomorphism positively influences HISI through subjective norms.

2.4 Digital health literacy as a moderator

Digital health literacy is conceptualized as an individual’s capacity to search for, access, comprehend, and critically appraise health information, as well as to apply the acquired knowledge to effectively address health-related issues. Within the digital context, DHL specifically denotes the proficiency to execute the aforementioned competencies in digital environments (16). The impact of GenAI’s perceived attributes on users’ intention to adopt the technology for health information seeking may be contingent on their level of digital health literacy. To date, no direct empirical evidence has been documented to confirm that DHL moderates the relationship between GenAI’s perceived characteristics and health information-seeking intention. Nevertheless, extant research indicates that individuals with higher DHL levels exhibit a stronger propensity to adopt generative artificial intelligence applications (23). Although the general public is increasingly reliant on GenAI for health-related services, users with low DHL cannot often screen health information critically and to comprehend or operationalize the technology effectively. Even if such users recognize the technical merits of GenAI, they may struggle to accurately evaluate the technology’s practical value, thereby attenuating the positive effect of GenAI’s perceived attributes on their usage intention. In contrast, a high level of DHL enables users to conduct a comprehensive and rational assessment of GenAI’s core attributes, which in turn amplifies their intention to use GenAI for health information-seeking purposes. In other words, among users with high digital health literacy, positive perceptions of GenAI’s technical characteristics are more likely to be translated into actual health information-seeking intention. Therefore, the following hypotheses are proposed:

H6a: Digital health literacy strengthens the relationship between PCP and HISI.

H6b: Digital health literacy strengthens the relationship between PCV and HISI.

H6c: Digital health literacy strengthens the relationship between PAN and HISI.

3 Data and methodology3.1 Sample and data collection

This online survey was conducted in December 2025, targeting individuals with a basic understanding of generative artificial intelligence (GenAI). Respondents were required to understand how to obtain health information through generative artificial intelligence and possess basic relevant knowledge. Participants who did not meet the criteria were excluded from the survey. Before the formal survey, the research team disseminated the survey link via the WeChat platform and recruited over 50 participants to complete a pilot study. Participants were instructed to fill out the draft questionnaire, assess each measurement item in terms of semantic coherence, logical consistency, and potential ambiguity, and provide constructive feedback for item revision. Based on the collected feedback, several measurement items were revised and refined to ensure the questionnaire’s logical rigor and readability. For the formal survey, Wenjuanxing,3 a professional online questionnaire platform, was adopted as the official distribution channel. This study adopts convenience sampling based on an online participant pool, which may lead to coverage bias. The sampling frame cannot fully represent the overall population of GenAI users in China, and potential undercoverage of certain subgroups is inevitable. To avoid duplicate submissions, the platform was set to permit only one response per participant via the exclusive survey link. In addition, all questionnaire items were designated as mandatory to eliminate missing values in the collected data. After rigorously excluding invalid questionnaires (e.g., incomplete responses, random answering), a total of 420 valid questionnaires were obtained, which served as the final sample for subsequent empirical analysis.

The demographic characteristics of the sample are presented in Table 1. The gender distribution was nearly balanced, with males accounting for 49.8% and females 50.2% of the participants. Some 34.3% of respondents were married, and nearly 58% fell within the 18–25 age group. Moreover, 22.9% possessed a bachelor’s degree or above, and 85.48% of participants self-rated their physical health as healthy or very healthy.

CharacteristicDemographicFrequencyPercentage (%)GenderFemale20949.8Male21150.2Age (years)18 and below102.418–2524357.926–4513732.646 and above307.1Educational levelJunior middle and below174.0Senior high153.6Junior college29269.5Bachelor degree and above9622.9Health statusVigorous9622.86Healthy26362.62Average5713.57Poor40.95Marital statusSingle27665.7Married14434.3Family income (monthly)￥5,000 and below6014.3￥5,000–￥10,00011627.6￥10,000–￥15,0009121.7￥15,000–￥20,0007618.1￥20,000 and above7718.3Frequency of useFrequent327.6Often9723.1Occasionally23054.8Hardly ever4210.0Never194.5

Demographic profile of respondents.

The number of samples is 420.

Source(s): Authors’ own work.

3.2 Measurements

The questionnaire comprised two sections: the first collected respondents’ basic demographic characteristics, while the second measured the core constructs of the study, namely perceived competence, perceived convenience, perceived anthropomorphism, trust in GenAI, subjective norms, digital health literacy, and GenAI-enabled health information-seeking intention.

All measurement items were adapted from established mature scales and appropriately modified to fit the research context. To ensure linguistic and cultural equivalence, a systematic translation and adaptation procedure was conducted. First, two bilingual researchers independently translated the original English scales into Chinese. Second, a third researcher synthesized the two translations into a preliminary Chinese version, which was then back-translated into English by two native English-speaking translators who were blind to the original items. The back-translated versions were compared with the original scales by the research team, and discrepancies were discussed and resolved iteratively until semantic equivalence was achieved. During this process, we conducted appropriate cultural adaptation. For instance, we adjusted the expressions to align with Chinese users’ habits of interacting with artificial intelligence and the practical usage context in China. Third, an expert panel comprising three scholars in health communication and one practicing physician reviewed the Chinese version for content validity and cultural appropriateness. Specifically, the perceived competence scale was sourced from Zhou et al. (40), perceived convenience from Qiu et al. (41) and Liang and Shi (42), perceived anthropomorphism from Priya and Sharma (43) and Al-Emran et al. (44), trust from Choudhury and Shamszare (5), subjective norms from Lai and Li (45), digital health literacy from Liu et al. (46) and Zhao et al. (47), and GenAI-enabled health information-seeking intention from Ling et al. (48). A 5-point Likert scale, anchored at 1 = strongly disagree and 5 = strongly agree, was employed to rate all items measuring the aforementioned constructs.

3.3 Technical analysis

Partial least squares structural equation modeling (PLS-SEM) was adopted as the primary analytical technique in this study, with its selection justified by three key considerations: First, PLS-SEM is particularly adept at analyzing complex models incorporating multiple constructs, mediating variables, and moderating variables—a critical advantage given the eight constructs included in the present study. Second, it does not impose stringent requirements for data normality, which enhances its flexibility for empirical research. Third, PLS-SEM boasts high statistical power and applies to both exploratory and confirmatory research, thereby yielding robust explanatory power for this study. Notably, the sample size of this study (n = 420) exceeds the minimum threshold of 205 samples recommended for PLS-SEM analyses. Accordingly, SmartPLS 4.0 was utilized to analyze the theoretical model, with a bootstrapping resampling procedure (5,000 resamples) employed to assess the statistical significance of path coefficients.

4 Data analysis and results4.1 Common method variance (CMV) test

Harman’s one-factor test was employed to assess common method variance (CMV) in the questionnaire data (49). The results revealed that the total variance explained by all extracted factors was 42.11%, with the variance accounted for by the first factor falling below the 50% critical threshold (50). This suggests that CMV did not constitute a major methodological concern in the present study.

4.2 Measurement model

Construct reliability was evaluated primarily via Cronbach’s alpha and composite reliability (CR). As presented in Table 2, all Cronbach’s alpha and CR values exceeded the 0.7 threshold, demonstrating satisfactory internal consistency of the measurement scales (51).

ConstructVIFItemsloadingCronbach’s CRAVEPerceived competence2.234I believe that GenAI can provide accurate health information.0.8800.8500.9090.7692.144I believe that GenAI is capable of accurately understanding my needs regarding health information.0.8821.916I am confident about the capabilities of GenAI.0.869Perceived convenience2.071I can use GenAI platforms to search for health information anytime and anywhere.0.8710.8710.9210.7952.572I find it convenient to search for health information using GenAI platforms.0.9042.43I can find the health information I need using GenAI platforms both quickly and easily.0.900Perceived anthropomorphism2.121GenAI tools are natural, and do not feel fake about them0.8780.8750.9230.82.469GenAI tools are conscious of their actions.0.8932.667GenAI tools are more humanlike and do not feel like machines.0.912User trust in GenAI2.553GenAI is trustworthy in terms of reliability and can provide reliable health information.0.9080.8530.9110.7732.599The health information provided by GenAI is honest and trustworthy.0.9071.727GenAI will not manipulate its responses, and making health-related decisions based on GenAI will not bring negative consequences to me.0.820Subjective norms2.693Most people I know believe that I should try using GenAI to seek health information.0.9050.8990.9370.8322.985Individuals around me with health concerns believe that it is necessary for me to attempt to use GenAI to seek health information.0.9232.723People I trust, such as family members or close friends, believe that I should attempt to utilize GenAI to seek health information.0.908Digital health literacy1.654I am able to comprehend health information found via the internet.0.8390.8160.8910.7311.873I am able to judge whether the health information found on internet is accurate or not.0.8491.988I am able to find information on the Internet to answer the questions on healthcare or disease treatment.0.877Health information seeking intention2.137The probability I would consider searching health information from the GenAI is high.0.8780.8730.9220.7972.354If I were to search for health information, I would consider searching from the GenAI.0.8892.668My willingness to search for health information from the GenAI is high.0.911

Reliability and validity tests of the constructs.

Source(s): Authors’ own work.

Convergent validity—defined as the degree to which measurement items reflect the same latent construct—is typically evaluated via the average variance extracted (AVE) and standardized factor loadings (52). The results indicated that all item standardized factor loadings exceeded the 0.7 threshold, satisfying the fundamental psychometric criterion. Additionally, all AVE values surpassed the 0.5 benchmark, confirming sufficient convergent validity for the constructs.

Discriminant validity was primarily examined using the heterotrait-monotrait ratio (HTMT) and the Fornell-Larcker criterion. As presented in Table 3, the square root of the AVE for each construct was greater than its correlation coefficients with all other constructs (53). Furthermore, the HTMT ratios for all construct pairs fell below the 0.90 cutoff value (54, 55). Collectively, these results demonstrated that all constructs in the model exhibited satisfactory discriminant validity.

1.2.3.4.5.6.7.1. Digital health literacy0.8550.6960.4750.5930.6180.6300.6012. Health information seeking intention0.823 [0.753, 0.887]0.8930.6250.6800.6850.7050.7003. Perceived anthropomorphism0.561 [0.463, 0.648]0.714 [0.651, 0.769]0.8940.6440.6340.6650.7044. Perceived competence0.709 [0.634, 0.777]0.787 [0.725, 0.841]0.745 [0.677, 0.802]0.8770.7340.6610.6945. Perceived convenience0.731 [0.649, 0.804]0.786 [0.734, 0.835]0.726 [0.661, 0.785]0.851 [0.801, 0.898]0.8920.6620.6526. Subjective norms0.733 [0.661, 0.797]0.795 [0.741, 0.845]0.749 [0.688, 0.806]0.754 [0.697, 0.807]0.747 [0.688, 0.801]0.9120.7287. User trust in GenAI0.716 [0.626, 0.797]0.808 [0.761, 0.853]0.816 [0.755, 0.867]0.807 [0.746, 0.862]0.747 [0.673, 0.815]0.831 [0.773, 0.883]0.879