Introduction
In today’s digital and AI-driven era, the way individuals seek health information is undergoing a gradual yet notable shift. With the rise of the millennial generation, primary reliance has moved from traditional news media to online search engines. Now, the rapid development of generative AI (GenAI) is beginning to change these behaviors once again, nudging users from occasional experimentation toward more purposeful use of AI tools for health-related searches. A recent study (Yun & Bickmore, 2025) found that while search engines and institutional health websites continue to be the dominant sources of online health information, about 21% of respondents had used large language model (LLM)-based chatbots for medical purposes. These GenAI tools are increasingly used to interpret symptoms and explore both preventive measures and treatment advice. Among those who reported using LLMs, nearly half (48.4%) said they followed the advice provided. In another national survey conducted in 2023, Mendel et al. (2025) found that 32.1% of U.S. adults reported using LLMs like ChatGPT for health-related queries, including interpreting symptoms, understanding treatment options, and reviewing test results. Usage was significantly higher among men and younger users (ages 18-29). These findings suggest that current LLM users are increasingly using GenAI tools for their health information seeking, which signal a growing openness to AI-assisted health behavior. Recent medical research has begun exploring the use of GenAI for delivering health information and education across various patient populations, including cancer patients (Alanezi, 2024), individuals undergoing total knee arthroplasty (Bains et al., 2024), and caregivers of pediatric cancer patients (Sezgin et al., 2025). These studies showed promising results in terms of information accuracy and perceived usefulness, while also cautioning that GenAI tools should complement professional medical guidance rather than replace it.
In addition, both news reports and peer-reviewed publications on GenAI continue to suggest that its capability for generating high-stake health information is steadily improving. For instance, Gilson et al. (2023) tested ChatGPT on tasks comparable to those found in medical licensing exams. Similarly Johnson et al. (2023) reported that ChatGPT’s accuracy and completeness in answering medical questions are comparable to the performance of licensed physicians. The advancement of LLMs in catching up with the cumulative medical knowledge humanity has amassed suggests that their use for medical and health-related information search may continue to grow as the technology evolves. Oeding et al. (2025) compared the trustworthiness and accuracy of ChatGPT-4’s responses about the Latarjet procedure with those retrieved using Google Search. For the top 10 patient FAQs, they found that ChatGPT-4’s answers were drawn entirely from academic sources (100%), while only 30% of Google’s results came from academic sources. In terms of clinical accuracy, ChatGPT-4 also outperformed Google, with a statistically higher accuracy score. These results are promising for the future use of LLMs in health information seeking.
These findings prompt an important question: what factors influence current LLM users to try and continue using GenAI for health information searches? While few studies have addressed this directly, existing research points to several key drivers. Zhou and Li (2024) offer insights into why users are shifting from traditional search engines to GenAI tools. Information overload was identified as a push factor driving users away from conventional search platforms. In contrast, information quality (Zhou & Li, 2024), accuracy of AI-generated content (Ghanem et al., 2024), or perceived credibility (Bains et al., 2024; Chevalier & Dosso, 2025) emerged as pull factors attracting users to GenAI. Previous research also suggest that risk factors such as privacy infringement and hallucination risks may inhibit users’ intentions to adopt GenAI tools (Lim et al., 2025), which could work as a mooring factor. This is because hallucination risk can discourage users from switching to or relying on GenAI tools instead of traditional search engines.
One of the salient risks in using GenAI is AI hallucinations—a phenomenon in which LLMs generate responses that appear plausible and factually grounded but are in fact misleading, inaccurate, or entirely fabricated (Lim et al., 2025; Nah et al., 2023; Rouzrokh et al., 2025; Shah, 2024). The presence of hallucinations in GenAI-generated medical information is well documented across recent studies (Colasacco & Born, 2024; Jin et al., 2023; Shah, 2024). Hallucinations concerns are also relevant in specialized medical information searches by patients and caregivers, where accuracy and up-to-date information are essential. For example, Sezgin et al. (2025) found that while widely used LLMs like ChatGPT performed well in expert evaluations for accuracy, clarity, and clinical usefulness, the responses often lacked proper source citations and sometimes included outdated information. This indicates that hallucinations remain a potential risk in medical information search and may discourage continued use of GenAI for seeking health and medical advice.
We assume that the ongoing improvement of LLMs and the introduction of new AI services by major tech companies will eventually lead to deeper integration of GenAI into everyday search behavior. In particular, the recent launch of Google’s AI Mode, a chatbot-style search feature that replaces traditional keyword-based search, marks a major milestone in the evolution of GenAI search (Mickle, 2025). Despite the growing adoption of GenAI tools and their potential to transform health and medical information searches, limited research has examined the factors that influence current users’ intentions to continue using these tools. To address this gap, the current study investigates how push and pull factors, along with concerns about AI hallucinations, relate to users’ behavioral intentions to continue using GenAI for health and medical information. We further note that prior studies have largely overlooked how users’ comparative evaluations of GenAI tools, particularly in relation to traditional web searches and medical professionals, could affect their continuance intentions. In light of this, a recent study found that users assess the usefulness of health information of GenAI-generated health information against more established sources, such as search engines and primary care physicians (Ayo-Ajibola et al., 2024).
To fill this research gap, the current research also examines how users assess the perceived usefulness of GenAI-based searches compared to conventional search engines and information obtained directly from physicians. To this end, we introduce the concept of comparative usefulness, grounded in Rogers’ (1995) diffusion of innovations theory. Drawing from Rogers’ (1995) notion of relative advantage, we define comparative usefulness as the perceived degree to which an innovation—here, GenAI tools—enhances a specific task (e.g., health information seeking) relative to other available sources, such as medical professionals or traditional online platforms. This construct captures how users weigh the benefits of GenAI tools, such as convenience, clarity, speed, or trustworthiness, against existing alternatives. Consistent with Rogers’ theory, it is the perceived advantage, not objective superiority, that better explains continued use and potential adoption.
This inquiry is grounded in the assumption that individuals in today’s multi-channel, fragmented media environment are unlikely to rely exclusively on a single source; instead, they are likely to use different types of media to complement, rather than replace, traditional sources (Dutta-Bergman, 2004a). This idea echoes the principle of relative constancy in media consumptions, which posits that overall time and spending devoted to media tends to remain stable (McCombs & Nolan, 1992). The channel complementarity theory (CCT) extends this idea by arguing that individuals allocate attention across media within specific content domains, such as health, sports, and politics, based on their need and involvement. Focusing on health information seeking, Rains and Ruppel (2016) found that users often select media sources that complement each other in convenience and tailorability. Guided by CCT (Dutta-Bergman, 2004a, 2004b), the current research explores whether users’ comparative assessments of GenAI search tools mediate the relationships between push (e.g., information overload), pull (e.g., credibility), and risk (e.g., hallucinations) factors and their intentions to continue using GenAI for health and medical information.
Theoretical Background
Channel Complementarity Theory: Expansion to AI
Channel Complementarity Theory (CCT) explains how individuals select and use media from among multiple competing sources to meet their specific needs and interests at a given time (Dutta-Bergman, 2004a). Drawing on the principle of relative constancy in media consumption(McCombs & Nolan, 1992), CCT challenges the displacement perspective in media consumption, which assumes that new media replace old ones. Instead, it adopts a motivational perspective, proposing that individuals actively use different types of media in parallel to fulfill their functional needs, rather than substituting one medium for another (Dutta-Bergman, 2004b).
While CCT is conceptually related to the relative constancy principle, it departs from it by focusing on content-specific media consumption (e.g., sports, health, politics) rather than total media consumption (Dutta-Bergman, 2004a). A few empirical studies (Moreno et al., 2023; Rains & Ruppel, 2016; Ruppel & Rains, 2012; Tian & Robinson, 2008) have applied CCT to health contexts to examine how individuals use competing media sources for health information seeking. Notably, Rains and Ruppel (2016) found that individuals tend to combine sources that are complementary in terms of tailorability and convenience, but are less likely to pair sources that differ widely in expertise and anonymity. Rains (2007) found that individuals’ trust in traditional health information sources, including healthcare providers, traditional media, and the internet, was positively associated with their likelihood of using the web to seek medical information. These findings support the view that individuals treat the internet as a complementary, rather than a substitutive channel. Tian and Robinson (2008) reported that greater attention to online health information was positively correlated with attention to traditional media sources, such as newspapers, magazines, radio, and television. They also found that more frequent visits to healthcare providers were associated with increased attention to both online and offline health information sources, which reinforces the complementarity hypothesis. Using national health survey data, Ruppel and Rains (2012) showed that individuals, when seeking health information, strategically combine multiple sources based on their unique functional strengths, such as access to medical expertise, anonymity, tailorability, and convenience.
CCT research has also been extended to examine how the credibility and trustworthiness of information sources influence information-seeking behavior (Lee et al., 2018; Moreno et al., 2023). Using two national surveys in Singapore, Lee et al. (2018) investigated how individuals’ trust in various media channels, including doctors, internet, and television, influenced their information seeking behavior related to cancer prevention. A cross-national survey-based study conducted during the COVID-19 pandemic found that citizens in Italy, Spain, and the United Kingdom engaged in multi-channel information-seeking, using complementary media sources rather than relying on a single outlet (Moreno et al., 2023). The study revealed that UK and Spanish citizens placed greater trust in local authorities and health professionals, while Italians exhibited higher levels of trust in national government institutions and official experts. Interestingly, negative perceptions of government communication were more prevalent among social media users.
As noted earlier, LLMs have emerged as credible and convenient complementary tools for health information seeking. In two U.S.-based surveys in 2023, Mendel et al. (2025) examined how adults use LLMS compared to search engines for health-related information seeking. They found that 95.6% of participants used search engines, while 32.1% used LLMs, with search engines being the most common first source (62%), compared to LLMs (14%). In addition, LLMs were perceived as less biased, less commercially influenced, and less likely to elicit negative emotions. A study by Ayo-Ajibola et al. (2024) found that the most common reason people used ChatGPT for health information was to determine whether visiting a health professional was necessary (47.4%), while the third most common reason was to clarify or verify information previously provided by a health professional (39.1%). These shifts reflect a growing reliance on AI not only for general medical knowledge but also for interpreting symptoms, evaluating treatment options, and even questioning professional diagnoses.
Zhou and Li (2024) found that users are increasingly switching from traditional search engines to GenAI tools for health information due to their fast delivery of perceived high-quality content in an easily digestible format. They also found that perceived anthropomorphism and interactive dialogue with LLM-based chatbots worked as pull factors to attract users while enhancing their engagement with content. As a result, many users now treat GenAI not as a replacement, but as a complementary extension to existing search engines. Sun et al. (2024) contended that GenAI’s constant accessibility, professional-level explanations, and conversational interface make it particularly appealing to digitally savvy users who regularly consult multiple online sources for health information. They found that trust in GenAI tools was significantly higher than in traditional search engines, especially for health-related tasks. Nevertheless, users in their study tended to see these tools as entry points for further exploration, rather than standalone solutions. Taken together, these finding suggest that current GenAI users perceive GenAI tools as complementary additions to their available health information sources. From the CCT perspective, GenAI at its current stage appears to function as a natural expansion of individuals’ information repertoires.
Perceived Information Overload in Online Health Information (OHI) Search
Perceived information overload (PIO) in OHI search refers to the perception that the amount of information is greater than “what the information seeker can manage effectively” and exceeds their “ability to process and handle the information” (Swar et al., 2017, pp. 416-419). Grounded in information processing theory, Swar et al. (2017) investigated how PIO affects users’ psychological well-being and their behavioral intention to continue searching for health information online. They found PIO negatively affected continuance intention indirectly through three dimensions of psychological ill-being (i.e., negative affect, depressive symptoms, and trait anger). In Soroya et al.’s (2021) research, social media use was positively associated with PIO, which in turn significantly predicted information anxiety. Zheng et al. (2023) contend that information overload often contributes to cyberchondria, a condition defined as the excessive or compulsive act of searching for health information online, which leads to anxiety rather than reassurance. For example, during the COVID-19 pandemic, individuals who searched for symptoms such as a cold or sore throat often came across alarming headlines about severe complications, which created unnecessary fear even when their symptoms were mild or unrelated (Soroya et al., 2021). This condition is associated with a psychological tendency to focus on worst-case scenarios, especially those linked to serious or life-threatening illnesses such as cancer or stroke. Using a three-wave panel survey in China, Zheng et al. (2023) reported that health-related searches through search engines increased perceptions of information overload, which, in turn, predicted higher levels of cyberchondria.
As major GenAI planforms are increasingly integrate search capabilities into their LLMs (Mickle, 2025), researchers are beginning to examine whether information overload associated with traditional search engines may act as a push factor that drives users to shift toward LLM-based tools. Proposing a push-pull-mooring (PPM) model, Zhou and Li (2024) found that Chinese users of both traditional search engines and GenAI tools often experience dissatisfaction with traditional search engines due to irrelevant or low-quality content, an overwhelming number of links, and conflicting advice (Zhou & Li, 2024). In contrast, GenAI tools provide concise, well-structured, and context-relevant responses through a conversational interface, which reduces users’ cognitive burden, increases the perceived value of information, and strengthens their intention to switch to GenAI tools for health information searches (Zhou & Li, 2024).
Treating PIO as a push factor that drives users away from traditional search engines during OHI searches, we predict that higher levels of PIO will be positively associated with GenAI users’ comparative evaluations of AI-assisted health searches. Specifically, we expect users experiencing PIO to view GenAI tools as more useful than both medical professionals (MDs) and other OHI sources such as WebMD or general search engines. Based on this rationale, we propose the following hypothesis:
AI Credibility in Online Information Search
Credibility remains a central factor in how users engage with and apply OHI. It influences not only user willingness to trust and follow medical advice but also their emotional responses, such as anxiety or reassurance. It is important because inaccurate or non-credible information can lead to misdiagnosis, unnecessary worry, or poor health decisions (Sbaffi & Rowley, 2017). While search engines have made health information more accessible, they have also introduced negative externalities. These include an overload of commercial content, clickbait headlines, information clutter, and sources that are biased, overly complex, or irrelevant (Sbaffi & Rowley, 2017). These challenges are particularly salient among older adults. Miller and Bell (2012) found that older adults reported significantly lower trust in OHI compared to younger users, and that trust was a stronger predictor of their likelihood to use the Internet for health information.
Since the use of GenAI for health-related information search is a relatively new trend, little research has examined how the perceived credibility of GenAI responses influences users’ intentions to continue using these tools. However, a few studies outside the health domain offer useful insights. For instance, Baek and Kim (2023) found that users’ trust in GenAI, specifically their perceptions of it as trustworthy, credible, and believable, was significantly and positively associated with their intention to continue using such tools.
An emerging question is how current GenAI users evaluate these tools in comparison to traditional OHI sources and healthcare professionals such as medical doctors. Two lines of research provide relevant insights. The first line of research involves studies that directly asked users to compare the usefulness of GenAI-based medical searches with that of medical doctors (MDs) and other OHI sources. For example, Ayo-Ajibola et al. (2024) found that 39.4% of respondents rated ChatGPT’s health information as better or much better than information from MDs, while 41.6% rated it as about the same and only 14.1% rated it as worse or much worse. When compared with other OHI sources, 63.2% of users rated ChatGPT’s information as better or much better, 24.3% rated it as about the same, and only 9.6% rated it as worse or much worse. These findings suggest that a substantial proportion of users perceive GenAI health information as not only credible but also comparatively more useful than both conventional sources and medical professionals.
The second line of research evaluated the information quality, accuracy, and reliability of GenAI tools using specific assessment criteria or by benchmarking them against standard measures such as medical doctor licensing exams. For instance, Walker et al. (2023) assessed ChatGPT-4’s responses to five high-burden hepato-pancreatico-biliary (HPB) conditions, comparing them against established clinical guidelines such as those from NICE and EASL. Although the median quality score (measured using the EQIP tool) was modest, the study found a 60% alignment rate between ChatGPT’s answers and guideline-based standards, with 100% internal consistency and substantial interrater agreement. They concluded that while ChatGPT’s content quality was comparable to static internet resources, it lacked contextual medical judgment, especially in complex, stage-dependent clinical situations. Johnson et al. (2023) evaluated ChatGPT-3.5 across 284 medical questions from 17 specialties and reported higher median accuracy and completeness scores. Most answers were rated between “mostly correct” and “completely correct,” with a median completeness score of 3, indicating responses were generally comprehensive. Notably, responses initially graded as inaccurate showed marked improvement when re-queried within two weeks, which implies ChatGPT’s strong potential for rapid model refinement. Together, these studies underscore both the promise and current limitations of GenAI in delivering reliable health information: while ChatGPT can often match physician-level performance in standardized contexts, its effectiveness still depends heavily on question type, clinical nuance, and the evolving nature of its training data. Taken together, these two lines of research support the idea that users’ perceptions of GenAI credibility may be positively related to their evaluation of GenAI’s usefulness in health information seeking. Thus, we propose the following hypothesis:
AI Hallucination Risk and Perceived Usefulness of GenAI Health Information
Despite the continual improvements in the accuracy and reliability of GenAI responses, a persistent and potentially unsolvable challenge remain. That is, assessing the quality of GenAI-generated content. Central to this issue is the phenomenon, in which LLMs generate responses that appear plausible and factually grounded but are in fact misleading, inaccurate, or entirely fabricated (Lim et al., 2025; Nah et al., 2023; Shah, 2024). Invoking the danger of AI-generated scientific research and writing, some critics opposed to calling it hallucinations and contended it is more accurately termed as fabrications or confabulation (Brender, 2023; Emsley, 2023). Some critics have challenged the use of the term “hallucinations” to describe AI-generated false information, especially in scientific research and writing. They argue that the term misrepresents the phenomenon and propose more precise alternatives such as “fabrications,” or “confabulations” (Brender, 2023; Emsley, 2023). Brender (2023) contended that confabulation, a term rooted in neuropsychology, better captures the nature of AI outputs that are logically structured but factually untrue, without implying malicious intent. Shiferaw et al. (2024), who evaluated the quality and risks of AI hallucinations in clinical responses related to chronic kidney disease and diabetes, concluded that ChatGPT should not be used as a standalone resource for high-stakes clinical decision-making, patient-specific care planning, or medication dosing. They emphasized that LLM tools should serve only as supplements to gold-standard clinical resources, and that all outputs must be reviewed by qualified domain experts.
Prior research suggests that users’ concerns about hallucination risk can negatively affect the perceived usefulness of GenAI tools across different contexts. For instance, Song et al. (2025) found that perceived hallucinations, in a crisis self-rescue scenario, significantly decreased users’ evaluation of AI service quality and satisfaction. In addition to generating factually inaccurate content, hallucinations may exacerbate cognitive biases by increasing users’ susceptibility to belief-consistent misinformation. Thus, perceived hallucination risks may have asymmetric effects depending on whether users perceive the content as belief-consistent or familiar, which has been shown to increase acceptance of false GenAI content (Shin, Koerber, et al., 2024). In a related study, Shin, Jitkajornwanich, et al. (2024) found that users were more likely to trust and accept AI-generated misinformation when it aligned with their prior beliefs or exhibited high familiarity.
Kim et al. (2025) found that when ChatGPT provided incorrect recommendations, users were less likely to rely on the GenAI tools and less inclined to act on its advice. In two U.S. surveys in 2023, Mendel et al. (2025) found that although users liked LLMs’ conversational and user-friendly interface, they still favored search engines for reliability and relevance. This preference may reflect ongoing concerns about medical accuracy, validation, and the potential for hallucinated or misleading information generated by LLMs.
Based on the review of potential negative aspects of GenAI tools for health-related information seeking, we posit the following hypothesis:
Comparative Evaluation GenAI’s Perceived Usefulness for Health Information
In the millennial era, the internet has become a primary source of health information, with over half of U.S. adults using search engines and social media for health-related purposes (Ayo-Ajibola et al., 2024). However, with the rise of GenAI, individuals’ health information-seeking behaviors are gradually shifting from search engines and social media platforms to GenAI tools. In the previous section, we introduced channel complementarity theory (CCT), which posits that individuals turn to multiple information channels not redundantly, but based on each channel’s unique informational or functional needs (Rains, 2007). At its core, CCT assumes that individuals do not necessarily replace one source with another, but instead build a complementary repertoire of sources that collectively complement each source’s limitations in areas such as expertise, accessibility, or tailorability. For instance, a patient may consult a doctor for diagnosis, an institutional site for official guidelines, and AI for quick clarification or simplified explanations. With the growing use of GenAI tools in health-related searches, there is increasing interest in how users perceive the usefulness of GenAI tools compared to traditional health information sources, such as medical professionals and institutional websites. Perceived usefulness refers to the degree to which a system enhances users’ task performance (Lim & Zhang, 2022) or decision-making quality (Davis, 1989; Doll et al., 1998). Doll et al. (1998) extended the application of this construct across various technological contexts. However, a potentially underexplored application is how users evaluate the comparative usefulness of emerging technologies like GenAI against previously dominant media or channels.
The importance of comparative usefulness can derive from Roger’s (1995) diffusion of innovations theory where he listed the notion of perceived relative advantage as the most important predictor in diffusion of innovation. According to Rogers (1995), relative advantage refers to the extent to which an innovation “is perceived as” (p.229) superior to the existing solution it replaces or supplements. Notably, Rogers (1995) emphasizes that the key determinant is not the innovation’s objective benefits, but the perceived advantage from the user’s standpoint. These advantages may include not only improved efficiency or accuracy but also social prestige, convenience, or personal satisfaction (Rogers, 1995). Thus, the construct of comparative usefulness captures users’ subjective evaluations of GenAI’s perceived relative advantage, which diffusion research consistently identifies as one of the strongest predictors of continued or widespread adoption.
Ayo-Ajibola et al. (2024) found that a significant proportion of users perceived GenAI-produced health information to be more useful than traditional sources. Specifically, 63.2% of respondents rated AI-generated information as better or much better than other online health sources, while only 9.6% rated it as worse or much worse. When compared with information from medical doctors (M.D.s), 39.4% considered the AI-generated content better or much better, while 41.6% viewed it as about the same. These results suggest that many users view AI as a comparably or even more useful tool than both institutional websites and healthcare professionals, particularly in terms of accessibility and perceived informational value. Kerstan et al. (2024) found that individuals having stronger comparative trust associations with AI than with physicians tend to exhibit a higher preference for AI over human doctors in healthcare contexts.
We contend that individuals evaluate comparative usefulness based on GenAI’s perceived functional benefits, such as its ability to reduce information overload, improve efficiency, and provide ease of use (Choudhury et al., 2025). This preference becomes particularly relevant when users seek quick, straightforward answers to general or non-urgent health questions (Shahsavar & Choudhury, 2023). In these situations, GenAI is often seen as more efficient than traditional search engines, which typically require users to sift through multiple links and continuously refine their queries. Xu et al.’s (2023) demonstrated this by comparing ChatGPT with Google. In their study, users who relied on ChatGPT completed search tasks in less time and reported lower cognitive effort. Their findings suggest that GenAI can streamline information retrieval by reducing the number of steps involved and eliminating the need to synthesize fragmented web content. Sun et al. (2024) also found that participants trusted ChatGPT more than Google for health information tasks, citing its interactive and concise response style as more effective for quick understanding. These results indicate that individuals’ perceptions of GenAI’s comparative usefulness reflect both its functional advantages and the degree of trust it holds relative to traditional health information sources.
Perceived Usefulness and Continuance Usage Intention of GenAI for Health Information Seeking
Continued usage intention refers to users’ willingness to maintain ongoing engagement with a technology beyond the initial adoption phase (Hong et al., 2006). A few researchers expanded the technology acceptance model (TAM) to examine continuance intentions of newly emerging AI technologies, including conversation AI (Ng, 2025), AI-based services (e.g., biometric systems) in the airport (Ku, 2025), AI painting apps (Yu et al., 2024), AI-powered banking apps (Lee et al., 2023), and text-based AI-powered customer service chatbots (Ashfaq et al., 2020). In health intervention research, understanding what factors motivate individuals to continue using new health technologies is essential because sustained user engagement is closely related to positive health outcomes (Wang et al., 2022). An important question is which factors influence continued use intentions of these health technologies. In a review of previous research on the studies focusing the continuance intentions of mobile health (mHealth), Wang et al. (2022) found that a few factors, including satisfaction, perceived usefulness, and trust, were significant predictors of continuance intention. A body of research showed that perceived usefulness is positively associated with continuance intentions. Zhang et al. (2018) found that the perceived usefulness of mobile health apps were positively associated with continued use intentions via satisfaction. Using a cross-sectional survey in the UK, Ukaegbu and Fan (2025) found that perceived usefulness was strongly correlated with continuance intention of COVID-19 contact tracing apps.
However, a key limitation in much of this research is that perceived usefulness is often measured based solely on the utility of a single technology in isolation. In reality, users typically assess the usefulness of a health technology by comparing it to other available options. For instance, the perceived value of a wearable glucose monitor may depend on how it performs in comparison to manual tracking or consultations with healthcare professionals. This comparative perspective is also relevant in the context of AI-driven health information seeking. When individuals use GenAI to find health or medical information, its perceived usefulness is often assessed in relation to that of human doctors and other OHI sources (Kerstan et al., 2024). Based on the well-established link between perceived usefulness and continuance intention, we propose the following hypothesis:
Mediating Role of Perceived Usefulness in AI-Driven Health Information Seeking
As reviewed earlier, several researchers have offered empirical evidence to assume potential relationship between perceived information overload, credibility, hallucination risk, and continued usage intentions of GenAI tools for health information seeking tools. We also proposed that these exogenous variables of perceived information overload, credibility, hallucination risk would be related to continuance intention, positively or negatively. Less known is whether perceived usefulness of GenAI searches, relative to medical doctor source and other information sources, may serve as mediating mechanism through which key factors like information overload, credibility, or hallucination risk influence their continuance intentions.
This prediction is based on previous research that has established perceived usefulness as one of the important factors to drive continued usage (Wang et al., 2022; Zhang et al., 2018).
The notion of channel complementarity suggests that individuals strategically use multiple sources to fulfill their health information needs, rather than relying on one channel to replace another (Rains & Ruppel, 2016; Ruppel & Rains, 2012). For instance, in the early age of digital and social media, users’ migration from offline to online channels for health information seeking was driven by perceived added value of new channels rather than their substitution to maximize the cross-channel synergy effect (Yang et al., 2013).
This strategic, non-substitutive approach to media use supports the idea that new information technologies become integrated into users’ existing repertoires based on their perceived functional value and relative benefits. Rather than replacing legacy media, search engines, or social platforms, users tend to evaluate whether GenAI adds value to their health information seeking, particularly when search engines cause information overload or when GenAI appears more credible than traditional sources.
Drawing on their experiences with health-related searches, users may judge GenAI’s usefulness by comparing it to the value they have found in web searches and medical advice. Building on this logic, the channel complementarity perspective suggests that the perceived usefulness of GenAI, relative to other sources, may explain how push factors such as overload, pull factors such as credibility, and a mooring factor such as hallucinations relate to users’ intention to continue using GenAI for health-related information seeking. Thus, we propose the following research question.
RQ1: How does the comparative usefulness of AI, relative to other health information sources (e.g., medical doctors and online platforms), mediate the relationship between perceived information overload, credibility, and AI hallucination and their continuance intention to use AI for health-related information seeking?
Method
After obtaining Institutional Review Board (IRB) approval from a large private university in South Korea, participants were drawn from Embrain’s national online panel of adults aged 19 to 69 and selected through a stratified cluster sampling method. The six major regions of South Korea—Seoul, Gyeonggi/Gangwon, Chungcheong, Honam, Gyeongsangbuk-do, and Gyeongsangnam-do—served as strata. Within each region, participants were randomly selected based on the proportion of the panel population in that region. A total of 1,200 respondents completed the survey, reflecting the demographic distribution of the Korean adult population. Panelists were invited via email and voluntarily participated. Those who completed the survey received a reward equivalent to KRW 3,000 in redeemable points for online or offline use. Of the participants, 49.3% were female and 50.7% were male, with a mean age of 46.25 years (SD = 13.03). Regarding educational attainment, 61.2% had completed a college program, 16.3% held a high school diploma, 11.9% had attended community college, and 10.6% had earned a four-year college degree or higher. Additionally, 23.1% of respondents reported an annual household income of $5,000 or more.
Questionnaire
We followed the back-translation procedure recommended for cross-cultural adaptation of research instruments (Sousa & Rojjanasrirat, 2011). The original questionnaire was initially developed in English. Two independent translators, both native Korean speakers fluent in English and Korean, participated in the translation process. One translator translated the original English version into Korean, and the second bilingual translator independently translated the Korean version back into English without access to the original. The research team then compared the back-translated English version with the original to identify and resolve any discrepancies through discussion, ensuring conceptual and linguistic equivalence.
Measures
Perceived information overload in health search
Respondents rated how frequently they experienced information overload while using search engines (e.g., Google, Naver) to find health-related information. This concept was measured using three items on a five-point Likert scale (1 = Never, 5 = Always), adapted from Zhou and Li (2024). A sample item includes “I am often distracted by the large amount of health information in search engine results” (M = 3.23, SD = .74, α = .74).
Credibility of GenAI’s health information
Perceived credibility of health-related information generated by GenAI tools was measured using a five-point semantic-differential scale. The measure included three bipolar adjective pairs: “Untrustworthy / Trustworthy,” “Unreliable / Reliable,” and “Unbelievable / Believable,” adapted from Ohanian (1990) (M = 3.37, SD = .72, α = .78).
Hallucination risk
Perceived AI hallucination risk was assessed using three items on a five-point Likert scale adapted from Lim et al. (2025). Respondents were asked how much they agreed with statements regarding the accuracy of AI-generated health information. A sample item includes, “I am skeptical about whether Generative AI outputs are based on facts or are fabricated” (M = 3.33, SD = .78, α = .82).
Perceived usefulness of AI compared with MD and other OHI sources
Perceived usefulness of AI compared to OHI sources was assessed using two items on a five-point Likert scale adapted from Ayo-Ajibola et al. (2024). These items measured the perceived usefulness of health information provided by AI relative to that from medical doctors and other OHI sources, such as WebMD, Mayo Clinic, government health websites (M = 3.05, SD = .69). Based on theoretical and methodological considerations, perceived usefulness of AI compared with other health information sources (COMPARE) was modeled as a formative construct rather than as a set of reflective indicators. This decision reflects the way individuals evaluate the usefulness of generative AI in real-world health information seeking. In practice, users do not assess AI-generated health information in isolation but in relation to multiple trusted sources. Medical professionals and reputable online platforms often serve as key benchmarks, and individuals compare the usefulness of AI-generated information against these diverse reference points. Therefore, combining these two items into a single formative construct more accurately captures the multidimensional and comparative nature of perceived usefulness. Following Hair et al. (2019), we first evaluated outer weights to assess each indicator’s contribution to the composite construct and then examined outer loadings as supplementary indicators. Both items, COMPARE1 and COMPARE2, yielded significant outer weights (β = .47 for MDs and β = .64 for other OHI, p < .001). These results demonstrated that both indicators contributed meaningfully to the composite variable COMPARE. In addition, both items exhibited high and significant outer loadings (λ= .92 for MDs and λ = .85 for OHI, p < .001), which were above the recommended threshold of .50. The VIF values for the two items were 1.549, and no multicollinearity concerns were identified.
Continuance intention
Respondents were asked how likely they are to continue using AI for general health information searches using three items on a five-point Likert scale adapted from Lim et al. (2025). A sample item includes, “I plan to continue using generative AI for health-related searches in the future.” (M = 3.55, SD = 1.10, α = .87).
Common methods variance (CMV)
We performed Harman’s one-factor test using SPSS to assess common methods variance (CMV) for the current study. Harman’s one-factor test is widely used method for detecting CMV under common conditions for survey research (Fuller et al., 2016). All items were entered into an exploratory factor analysis using principal axis factoring without rotation. The results showed that the variance explained by the single factor was 30.12%, which is below the 50% threshold (Podsakoff & Organ, 1986). This suggests that common method bias is not a significant concern in this study.
Measurement model
We assessed the measurement model using partial least squares structural equation modeling (PLS-SEM) via SmartPLS 4. PLS-SEM was selected based on two methodological considerations consistent with our study objectives and data characteristics. First, because the primary aim of this research was prediction and theory development, PLS-SEM was more appropriate than covariance-based SEM (CB-SEM), which is better suited for theory testing and confirmation (Hair et al., 2019). Second, the model included both reflective and formative constructs. PLS-SEM accommodates formative indicators without requiring the strict identification conditions that CB-SEM typically demands. Therefore, this approach can overcome the convergence and estimation issues that often occur in CB-SEM when formative constructs are included (Hair et al., 2019).
We evaluated both convergent and discriminant validity for each latent construct. Table 1 presents the factor loadings for all reflective constructs, along with Cronbach’s alpha, composite reliability (CR), and average variance extracted (AVE). All standardized factor loadings exceeded the recommended threshold of .60 (Falk & Miller, 1992), which suggest that observed indicators had strong relationships with their corresponding latent constructs. The composite reliability (CR) values ranged from .85 to .92, which surpassed the .70 threshold recommended by Nunnally and Bernstein (1994). The average variance extracted (AVE) for each construct exceeded the recommended threshold of .50 (Fornell & Larcker, 1981), with values ranging from .74 to .86. These results confirmed that all items effectively converged on their respective constructs and indicated a high internal consistency and reliability of the items within each construct. Table 1 also presents the factor weights for the two formative constructs.
To assess discriminant validity, we applied the Heterotrait-Monotrait (HTMT) ratio of correlations, which offers greater reliability than the traditional Fornell-Larcker criterion in variance-based SEM (Henseler et al., 2015). As shown in Table 2, all HTMT values ranged from .04 to .56, well below the commonly accepted threshold of .85. These results indicate all measured constructs established discriminant validity. For model fit evaluation, we used the standardized root mean square residual (SRMR), an approximate fit index recommended for PLS path modeling (Henseler et al., 2016). The SRMR value for the estimated model was .05, which falls below the recommended cutoff value of .08 (Henseler et al., 2016; Hu & Bentler, 1999).
Results
Structural model
To test the proposed hypotheses, we employed Partial Least Squares Structural Equation Modeling (PLS-SEM) with mediation analysis using SmartPLS 4. We assessed the significance of the path coefficients using bias-corrected and accelerated bootstrapping with 10,000 bootstrap resamples and reported 95% confidence intervals (CI). The estimated standardized coefficients for both direct and indirect paths are illustrated in Figure 1 and summarized in Table 3.
Hypothesis testing
Table 3 presents a summary of the hypothesis testing results. In H1a, we predicted that perceived information overload would be positively associated with perceived comparative usefulness of GenAI, relative to MDs and OHI sources. H1b posited a positive association between perceived credibility of GenAI and its perceived comparative usefulness. H1c hypothesized that perceived AI hallucination risk would be negatively associated comparative usefulness. The results supported all three hypotheses: perceived information overload (β = .16, p < .001) and credibility (β = .45, p < .001) were both positively associated with perceived comparative usefulness, whereas hallucination risk was negatively associated (β = −.18, p < .001).
H2 posited that perceived comparative usefulness of GenAI would be positively associated with continuance usage intention. The results confirmed this relationship (β = .29, p < .001), which supports H2.
To answer RQ1, we ran a mediation analysis. The results of mediation analysis showed that perceived comparative usefulness significantly mediated the relationships between all three exogenous variables and the endogenous variable of continuance usage intentions. More specifically, perceived information overload (β = .05, 95% CI [.03, .07]) and AI credibility (β = .13, 95% CI [.10, .17]) exhibited significant positive indirect effects, while perceived hallucination risk showed a significant negative indirect effect on continence intention to use GenAI for health information seeking (β = −.05, 95% CI [−.08, −.03]) through comparative usefulness.
Discussion
Grounded in the push-pull-mooring (PPM) framework, this study examined how three factors—perceived information overload from search engines (push factor), perceived credibility of GenAI (pull factor), and perceived hallucination risk of GenAI (mooring factor)—are associated with users’ perceptions of GenAI’s usefulness relative to medical professionals and other OHI sources. We then tested whether this perceived comparative usefulness is positively associated with users’ intention to continue using GenAI for health information seeking. Finally, we explored whether comparative usefulness mediates the relationships between the three perceptions and continuance intention.
Consistent with our predictions, we found that the frequency of experiencing information overload while using search engines for health information and the perceived credibility of health-related information generated by GenAI were positively associated with the perceived comparative usefulness of GenAI. This finding aligns with Choung et al. (2023), who demonstrated that users’ trust in AI agents strengthened their perceptions of usefulness, especially when those agents were perceived to offer clear, relevant, and timely assistance during health-related decision-making tasks. We also confirmed a negative association between AI hallucination risk and comparative usefulness. Regarding effect size, we observed the strongest path coefficient from perceived credibility to comparative usefulness. This result suggests that users’ evaluations of GenAI’s usefulness relative to MDs or other OHI sources are strongly influenced by perceived reliability and accuracy of LLM-based health information. Understanding the relative influence of push and pull factors is important because users’ adoption and continued use of GenAI tools may vary depending on how useful they perceive these tools to be. For example, “optimizers” may perceive greater comparative usefulness in GenAI tools for health information seeking, while “skeptics” may place more weight on hallucination risks (see Lim et al., 2023).
These findings also support the push-pull-mooring (PPM) framework, widely used in migration and information systems research to explain why individuals switch from one technology or system to another (Hsieh et al., 2012). As applied to health-related information seeking via GenAI, information overload appears to function as a push factor, potentially driving users away from traditional search engines. Meanwhile, the perceived credibility of GenAI acts as a pull factor that encourages continued use. Hallucination risk serves as a mooring factor that may inhibit users’ continuance usage intention. It is important to note that our model tested a parsimonious set of factors, with one variable representing each of the push, pull, and mooring categories. Future research should expand this framework by including a broader range of factors (see Zhou & Li, 2024).
Our research revealed a positive association between the perceived comparative usefulness of GenAI and users’ intention to continue using it. This finding is consistent with prior research showing that perceived usefulness is an important predictor of continued engagement (Wang et al., 2022; Zhang et al., 2018). This finding also aligns with Kerstan et al.’s (2024) research that found that individuals perceiving AI as more trustworthy and competent than human doctors were more likely to prefer AI-led health services. A notable contribution of the current research lies in its focus on relative usefulness. Unlike other studies that measured usefulness as an independent construct, typically evaluating a single platform or application, this study assessed it in comparison with two other available sources (i.e., medical doctor and other online health information sources) (cf. Ayo-Ajibola et al., 2024). To answer RQ1, we examined whether the perceived comparative usefulness of GenAI (COMPARE) mediated the relationship between three key predictors—information overload, perceived credibility, and hallucination risk—and users’ continuance intention (CI). All three variables demonstrated statistically significant, albeit modest, indirect effects on CI through COMPARE. Among them, the mediation effect was strongest for perceived credibility. This result suggests that users who found GenAI more credible also tended to perceive it as comparatively more useful, which in turn was associated with greater intention to continue using it. In contrast, the mediation pathway from information overload to CI through COMPARE was weaker but still significant. Notably, the direct path from perceived hallucination risk to CI was non-significant; however, its indirect effect through COMPARE was significantly negative. This indicates that when users perceived a higher risk of hallucinations, their evaluation of GenAI’s comparative usefulness diminished, which then corresponded with a lower intention to continue its use. These results suggest that users’ assessment of GenAI’s comparative usefulness serves as a mechanism through which their perceptions of credibility, information overload, and hallucination risk are related to their intentions to continue using it.
Theoretical and Practical Implications
The current research has identified push, pull, and mooring factors influencing LLM users’ continued use of GenAI for OHI search. As expected, our findings confirm that information overload drives users away from traditional search engines and toward GenAI platforms. The results also suggest that perceived credibility of LLM-generated health information could significantly attract and retain users, which serve as a pull factor. This aligns with Bains et al.’s (2024) research that showed that patients recovering from knee surgery were more likely to prefer ChatGPT’s responses over those provided by nurses when the content was perceived as credible. Their study showed that when GenAI-generated content is perceived as credible and reassuring, patients are more inclined to trust and adopt GenAI tools in healthcare services. By showing a positive correlation between credibility and continued health search using GenAI tools, the result also supports Lee et al.’s (2018) comprehensive model of information seeking, which extended CCT by incorporating trust as “information-carrier factor” in predicting health information seeking behavior. The finding that credibility is positively associated with the formative construct of comparative usefulness of AI search, relative to both medical doctors and other OHI sources, also supports Lee et al.’s (2018) notion of trust complementarity. According to this view, individuals tend to develop small clusters of trusted sources that work together to meet their information needs, rather than relying on a single dominant source. It is also notable that perceived credibility was positively linked to the comparative usefulness of GenAI-assisted health searches relative to both MDs and other OHI sources. This suggests that the more credible users perceive GenAI-assisted searches to be, the more useful they find them compared to traditional sources. Future research could further explore whether these perceptions and comparative evaluations differ across age groups, such as between older and younger adults.
We also found that AI hallucinations lowered comparable usefulness of GenAI, but this exogenous variable did not have a statistically significant impact on continuance intentions. This finding is consistent with recent evidence suggesting that hallucination risk does not strongly deter continued use of GenAI for various contexts. For instance, Lim et al. (2025) found a statistically significant positive relationship between perceived hallucination risk and continuance intention, which suggests that many users continue using GenAI tools despite recognizing their potential for factual errors. This counterintuitive pattern may reflect users’ growing familiarity with GenAI’s limitations and a willingness to tolerate minor inaccuracies in exchange for convenience or speed. Ghanem et al. (2024) offer a possible explanation for this tolerance: in their evaluation of ChatGPT’s responses to frequently asked questions about osteoporosis, the AI achieved an average accuracy score of 91%, with no responses rated as inaccurate or harmful. Most answers were considered “excellent” or “accurate with minimal clarification.” These findings suggest that, at least in domains with well-established medical knowledge, current GenAI users may perceive the actual risk of being misinformed as minimal. Together, these results help explain why hallucination risk may not serve as a major deterrent—users either don’t frequently encounter problematic content, or they feel confident enough to assess its reliability when they do.
Findings from the current research suggest a promising outlook for the growing role of GenAI models in the OHI search space, which has long been dominated by Google search engines. Without question, LLMs appear poised to capture a larger share of this market due to speed and generally accurate responses. However, this research also suggests that LLM-based health information search must prioritize delivering verifiable and up-to-date information. Sezgin et al. (2025) warned that LLMs should not be viewed as replacements for clinical advice. They further emphasized that to improve reliability and reduce the risk of hallucinations, LLMs should be enhanced with retrieval-augmented generation (RAG) methods that allow integration of verified, real-time sources into responses. From the development perspective, Rouzrokh et al. (2025) also note that even with RAG methods, additional safeguards, such as confidence indicators, are needed to help users better judge how much to trust each response. Although such technical solutions lie outside of users’ control, understanding their importance may help users critically evaluate GenAI outputs and avoid overreliance.
We also believe that our findings carry practical implications for both public health platforms and AI system designers. While the results suggest that GenAI can effectively reduce information overload in health information searches, public health platforms are still likely to remain trusted sources, especially given ongoing concerns about GenAI’s limitations, including hallucinations and lack of institutional authority. In this context, collaborations between authoritative public health organizations (such as the Centers for Disease Control and Prevention or national health departments) and major AI developers could yield meaningful synergies. Specifically, integrating validated public health content into GenAI systems may enhance both efficiency and perceived credibility. Even prior to the rise of GenAI tools, during the COVID-19 pandemic, the World Health Organization deployed an AI-powered chatbot via WhatsApp to deliver real-time, trustworthy health information, which demonstrated the potential of AI in expanding public health communication (Panteli et al., 2025). To support such efforts, AI system designers should prioritize interoperability with public health databases and embed transparency-enhancing features such as source citations, trust indicators, and real-time content verification (Sun et al., 2024).
Limitations and Suggestions for Future Research
To enhance the generalizability of the findings, this study employed a probability-based stratified cluster sampling method. Despite this strength, the study has several limitations commonly associated with single-time-point cross-sectional surveys. First, the user perceptions reported in this study reflect views at the time the survey was conducted. However, as LLMs continue to be developed by diverse startups and global companies, tailored to specific populations, and integrated into traditional search engines, user perceptions and behaviors may shift in response to these evolving technological affordances. Subsequent studies may benefit from a longitudinal survey design to further validate and expand upon these findings.
Second, this study focused solely on continuance usage intention as the endogenous variable. Given the relatively early stage of user adoption of GenAI for health information seeking, future research should consider other important engagement behaviors including switch, initial adoption, and discontinuance intentions.
Third, the construct of perceived comparative usefulness was operationalized as a formative construct using two reference points—medical professionals and other OHI sources. While this approach reflects the multidimensional nature of comparative evaluation, it may limit the generalizability of the construct’s psychometric properties. A promising direction for future research involves modeling comparative usefulness as a reflective latent variable, informed by more comprehensive scales developed in prior literature, to allow for deeper construct validation and replication across contexts.
Fourth, this study examined general perceptions of the comparative usefulness of GenAI without investigating potential subgroup differences based on demographic factors (e.g., generational cohorts such as Gen Z vs. millennials) or psychological traits (e.g., digital literacy). However, such individual differences may serve as important moderators in how users assess and interact with AI-generated health information. For example, younger users who are more digitally immersed may be more inclined to view GenAI as a useful source of health information, while older individuals may place greater trust in traditional sources such as medical professionals or established health websites. Likewise, individuals with higher levels of digital literacy may be better equipped to critically evaluate the credibility and utility of AI-generated health information, which could positively influence their perceived usefulness of GenAI and their intention to continue using it. Future research should systematically examine these subgroup variations to better understand how demographic and psychological characteristics shape users’ evaluations and behavioral responses to GenAI in health contexts.
Finally, this study was conducted in South Korea, where limited research exists on the use of GenAI for health information seeking. Although major GenAI tools support the Korean language, little is known about the quality and accuracy of health information generated in Korean, which may differ from outputs in English. Future studies should include user evaluations of GenAI response quality in their native language and account for demographic and contextual variables to better understand cross-cultural differences in how GenAI is used for health-related information seeking.
Conclusions
Despite its limitations, this study contributes to a growing body of research on GenAI use in health information seeking by applying CCT and the push-pull-mooring model. It examined how comparative usefulness mediates the relationship between perceived overload, credibility, hallucination risk, and continuance intention. The findings offer empirical support for understanding what drives continued use of GenAI for health information seeking, which is critical for evaluating its long-term role and integration into users’ health information behaviors.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
Print
