Abstract
Introduction
The integration of artificial intelligence (AI) into health information seeking is transforming health promotion. Understanding how users accept and trust these communication technologies is critical for health communication and cancer control. This study examined how the Technology Acceptance Model II (TAM II) applies to colorectal cancer information seeking, comparing link-based search (e.g., Google search) versus generative response paradigms (e.g., ChatGPT/AI) while examining trust, perceived threat, and contextual factors in technology use decisions.
Methods
A prospective, randomized 2×2 factorial experiment was conducted with 764 Texas adults randomly assigned to conditions to view either Google search results or ChatGPT responses for colorectal cancer symptoms, presented in either high-concern or low-concern scenarios. Participants completed validated measures including TAM II constructs adapted from Davis (1989) and Kamal et al (2020), multidimensional trust scales, Extended Parallel Process Model threat measures (Witte, 1992), and technology-related stress items, all demonstrating acceptable reliability (α > .77). Data analysis included two-way ANOVAs, correlation analysis, and stepwise regression modeling.
Results
Google search received significantly higher ratings than AI across all Technology Acceptance Model II constructs. Technology preferences appeared to reflect multiple factors including interface familiarity, trust in information sources, and usability expectations, with traditional search benefiting from established user mental models and transparent source attribution. Trust emerged as the strongest predictor of behavioral intention. No significant main effects were found for concern level, and no interaction effects emerged between technology type and concern level, indicating that technology preferences remained consistent regardless of symptom severity.
Conclusions
For cancer control and prevention, these findings suggest that patients seeking colorectal cancer symptom information may be more likely to trust and act upon traditional search results than AI-generated responses, focusing on technology use intentions for health information seeking that directly inform cancer screening and care-seeking behaviors, potentially affecting screening behaviors and care-seeking timing. Current AI implementations may not optimally serve health information needs with lower acceptance potentially related to limited source transparency and increased cognitive demands compared to familiar search interfaces, as suggested by preference patterns. Cancer control professionals should anticipate that the growing integration of AI into health information seeking may influence the public’s cancer symptom evaluation and screening behaviors.
Plain Language Summary
When people have worrying symptoms like stomach pain or changes in bowel habits, they often search online for information before deciding whether to see a doctor. Today, people can get health information from traditional search engines like Google or from newer artificial intelligence (AI) chatbots like ChatGPT. But we don't know which type of search people trust more, especially when they're worried about serious health problems like colorectal cancer. We studied 762 adults in Texas to find out how people feel about using AI versus Google for health information. We showed participants realistic examples of both Google search results and AI responses about colorectal cancer symptoms. Some people saw information about serious symptoms (like blood in stool), while others saw less worrying symptoms (like occasional gas). We then asked them how useful, easy to use, and trustworthy they found each type of search, and whether they would actually use it for health questions.Our results showed that people strongly preferred Google over AI for health information across every measure we tested. They found Google more trustworthy, easier to use, and more helpful. Surprisingly, it didn't matter whether the symptoms were serious or minor - people consistently chose Google over AI regardless. The biggest factor in whether people would use a search method was how much they trusted it.These findings matter because AI is becoming more common in health information, but people may not be ready to trust it yet. For cancer prevention, this means patients might be more likely to act on information from familiar sources like Google rather than AI. Healthcare providers should be aware that patients may have different levels of trust in AI-generated health advice, which could affect how they interpret symptoms and decide when to seek medical care.
Keywords
Introduction
Health-information seeking is one of the most common online activities and is undergoing dramatic transformation as artificial intelligence (AI) becomes more integrated into user experiences.1,2 This transformation is particularly consequential for cancer-related information seeking, where individuals may search for symptom information, screening guidance, or treatment options that directly influence life-saving healthcare decisions. Large language models (LLMs) are fundamentally altering health-search behaviors and clinical conversations focused on patient education, representing a paradigm change in how individuals access, process, and act upon health information. 3 Traditional search engines like Google have long served as primary gateways to health information, presenting users with ranked lists of links to websites, medical databases, and institutional resources. AI-powered interfaces offer a fundamentally different interaction model where users can receive what seem like settled summaries of advice or engage in dialogue-like exchanges of seemingly contextual responses to health queries.4-6 The platforms represent fundamentally different interaction paradigms: traditional search engines present users with multiple sources requiring evaluation and synthesis while AI systems provide pre-synthesized responses that alter the relationship between user and information. The widespread integration of artificial intelligence into health-information seeking poses immediate risks and opportunities for public health outcomes as well as individuals’ health as the consumers of AI-generated search results.
A central challenge is transparency. AI-generated content now appears within familiar search environments, often with limited disclosure, 7 shifting users from evaluating multiple sources to relying on synthesized outputs. This shift can alter risk appraisal, symptom interpretation, and care-seeking, with potential downstream consequences, such as delayed help-seeking if AI responses provide false reassurance or heightened anxiety if outputs emphasize worst-case scenarios without clinical context.
Along with deliberate interactions with conversational AI platforms like ChatGPT, AI-generated health information is also increasingly embedded within traditional search interfaces with varying levels of disclosure. 8 Recent research on online health information systems highlights transparency challenges, noting that “the frequency and conditions under which pre-screening or blocking [of content] occurs are uncertain” and that AI systems may produce results without employing official channels or vetting by experts. 9 Studies examining Google’s AI Overviews feature demonstrate how “AI generates answers without human guidance” and relies on “predictive text generated by large language models trained on information gathered from all over the web” including unvetted sources that can influence medical information without users’ full awareness. 10 This incorporation of unvetted sources is particularly concerning because medical misinformation can directly influence care-seeking behaviors, symptom interpretation, and attitudes toward medical experiences.
In addition to the search experience, the degree of concern people experience around specific health issues can influence health information-evaluation strategies and subsequent health behaviors. Research in health psychology demonstrates that perceived health threat significantly influences information processing strategies, with individuals under higher threat showing both increased motivation to seek information and heightened scrutiny of information quality.11-13
While threat increases motivation to seek information generally, it may simultaneously increase preference for familiar, trusted information sources over novel technologies when evaluating high-stakes health decisions, as individuals can seek to minimize additional uncertainty when facing health-related stress. Under stress they may view unfamiliar technologies as introducing risk into critical decisions while for low concern health topics, users may be more willing to rely on novel interfaces, consistent with technology acceptance research demonstrating that convenience becomes a more influential adoption factor in lower-stakes contexts. 14 Understanding how concern level moderates preferences between AI and traditional search technologies is therefore essential for predicting adoption patterns across the spectrum of health-information seeking options.
This research examines how users evaluate and respond to AI search versus traditional search results for colorectal-cancer-information seeking, applying an extended Technology Acceptance Model II (TAM) framework to identify the psychological factors that predict technology acceptance in high-stakes health contexts. The findings provide essential insights for understanding how embedded AI systems reshaping health-information consumption patterns are while offering guidance for cancer control practitioners, healthcare providers, and technology developers navigating this rapidly evolving landscape. To examine these dynamics, this study employs a 2×2 factorial design comparing presentations of newer, AI search results versus a known search engine format across high- and low-concern colorectal cancer (CRC) information-seeking scenarios, allowing identification of how health urgency moderates technology acceptance patterns.
Artificial Intelligence and Health Information
Embedded AI systems may improve accessibility for those facing language barriers and enhanced support for people with varying levels of health literacy.15-17 Research also demonstrates AI chatbots’ effectiveness in promoting health behavioral changes through personalized communication. 18 AI systems offer potential advantages including personalized health communication, improved accessibility for users with limited health literacy, conversational interfaces that may reduce barriers for sensitive topics, and capacity for real-time adaptation to individual user needs and preferences. These benefits are counterbalanced by challenges related to information quality, algorithmic bias, and the risk of AI systems generating misleading or inappropriate medical guidance without users recognizing the source or limitations of the information.8,19 The behavioral impact of health-information seeking extends beyond knowledge acquisition, often also serving as a coping mechanism for managing health anxiety, preparing for medical encounters, and influencing behavioral choices in domains such as mental health and dietetics.20-23
When individuals search for health information, they engage in complex processes involving information gathering, meaning-making, risk assessment, and emotional regulation. The introduction of AI-generated content into search experiences can alter these dynamics, potentially affecting how individuals perceive health risks, understand medical conditions, and interact with healthcare providers.24,25 Unlike traditional search results that present multiple sources requiring comparison and evaluation, AI systems provide synthesized responses that may reduce users’ engagement in deliberative information processing, potentially decreasing critical assessment despite offering greater convenience and apparent comprehensiveness. 24 The transparency challenge is particularly acute when AI-generated responses appear within traditional search experiences without explicit disclosure, creating a heterogeneous landscape where different users apply different cognitive and evaluative frameworks, potentially leading to divergent health behaviors and outcomes.1,26-29
Given these transformations, research must address how users interpret and respond to AI-generated content within familiar search interfaces. This may be of particular relevance to healthcare providers, as they may increasingly encounter patients whose medical knowledge and expectations have been shaped by AI-synthesized information, requiring adaptation in communication strategies and patient education approaches.28,30 During more than two decades since the launch of modern search engines, as many users gained experience with online health information seeking, they developed evaluation strategies including source credibility assessment and multi-source comparison. 31 The integration of AI-generated content disrupts established evaluation frameworks, potentially leaving individuals without effective strategies for assessing information quality.
Colorectal Cancer as a Critical Context
Colorectal cancer (CRC) information seeking provides an optimal context for examining AI versus traditional search acceptance because early symptoms are ambiguous, leading to repeated cycles of information-seeking and iterative judgments about whether medical attention is warranted. This leads to decision processes where information quality, source consistency, and emotional framing can directly impact healthcare use.32,33 Colorectal cancer ranks as the second leading cause of cancer death in the United States, with an estimated 107,320 new cases of colon cancer and 46,950 new cases of rectal cancer expected in 2025. Also, age demographics are shifting dramatically, with incidence rates increasing by 1-2% annually among adults younger than 50 since the mid-2000s. 34
Early CRC symptoms—changes in bowel habits, blood in stool, persistent abdominal discomfort, and unexplained fatigue—are often subtle, intermittent, or easily attributed to benign conditions.35-37 This ambiguity leads to repeated cycles of online information seeking before individuals decide whether to seek medical care. In these scenarios, individuals often rely on online information to make sense of symptom uncertainty in the absence of clinical guidance, making the quality, clarity, and framing of available health information particularly influential. 38 Information that minimizes risk or normalizes symptoms may reassure users but delay care-seeking, whereas information that highlights the potential seriousness or clear action may increase perceived threat and encourage medical consultation. Accordingly, the quality and framing of online health information can significantly shape symptom interpretation and subsequent healthcare decisions.
The screening landscape adds further complexity to the importance of information-seeking. Despite screening effectiveness in reducing CRC incidence and mortality, screening rates remain low with notable disparities. In Texas, where this study was conducted, only 36% of patients in Federally Qualified Health Centers were up-to-date with screening in 2021, significantly below both the national FQHC screening rate of 42% and the Healthy People 2030 goal of 74%.39-41 Understanding how different search technologies influence individuals’ perceptions of CRC risk, screening necessity, behavioral intentions, and symptom significance becomes crucial for developing effective public health interventions in an era where users may unknowingly rely on algorithmically-mediated health information.
Technology Acceptance Model II and Extensions for AI Health Information
The Technology Acceptance Model II (TAM II) provides a robust theoretical framework for understanding technology adoption in health-related behaviors, identifying perceived usefulness, perceived ease of use, and subjective norms as among the key factors influencing attitudes and behavioral intentions.42,43 Hypothesis 1. Perceived usefulness of AI-powered health search will be positively related to intention to use AI for health information seeking. Hypothesis 2. Perceived ease of use of AI-powered health search will be positively related to intention to use AI for health information seeking.
Originally developed by Davis
44
and refined by Venkatesh and Davis,
42
TAM II has demonstrated stability across technological transitions from early internet adoption to mobile health applications and AI-powered health information systems.
45
Recent research confirms these core TAM constructs while identifying additional psychosocial barriers that may be particularly important for healthcare technology adoption.
46
The model’s applicability to health information seeking stems from its recognition that technology adoption involves complex interactions between functional benefits, usability perceptions, and social influences, all factors that become particularly critical in health contexts where information quality and uncertainty can have significant implications. Figure 1 presents the theoretical framework examining known TAM II relationships in addition to trust and threat extensions described below. Extended technology acceptance model II for colorectal cancer information seeking: Effects of concern level, trust, and perceived threat
The unique characteristics of AI-generated health information, particularly when embedded within conventional search interfaces, offer opportunities for theoretical extensions beyond traditional TAM II constructs. For example, Kleine et al 47 showed that healthcare professionals’ intentions to use AI diagnostic tools were significantly predicted by perceptions of AI usefulness and confidence in operating AI systems effectively while Woodcock et al 48 found similar patterns for consumer-facing AI health applications with additional considerations for transparency, trust, and personalization.
Trust as Central Extension: Trust represents perhaps the most critical addition to TAM II for AI health applications because unlike conventional technologies where users evaluate system outputs through long-established criteria, AI systems are new enough to require users to place confidence in opaque algorithmic processes. This study conceptualizes trust multidimensionally, incorporating both institutional trust (confidence in the AI provider organization) and situational trust (confidence in specific AI-generated responses). Research suggests AI systems face higher trust thresholds than traditional technologies, particularly in healthcare contexts where trust can encompass both epistemic dimensions (confidence in information accuracy and reliability) and procedural dimensions (confidence in system processes and transparency).
27
These trust mechanisms may be compromised when AI-generated content appears without clear disclosure within routine search experiences or when users feel uncertain or unfamiliar with the technology. Hypothesis 3. Trust in AI-generated health information will moderate the relationship between perceived usefulness and behavioral intention, such that the positive relationship between usefulness and intention will be stronger when trust is high compared to when trust is low. Hypothesis 4. Trust will be a stronger predictor of behavioral intention for AI-powered search compared to traditional search engines.
Perceived Threat and Stress as Contextual Factors
Health information seeking often occurs under uncertainty (i.e., ambiguous symptom interpretation, unclear health risk) and anxiety conditions (e.g., worry about potential serious illness) that may significantly influence how individuals evaluate AI versus more familiar search technologies. We incorporate perceived threat and technology-related stress to capture users’ assessments of potential negative consequences from relying on different information sources and the cognitive-emotional burden of using these technologies for health decisions.
13
The Extended Parallel Process Model (EPPM) suggests threat perceptions (conceptualized with susceptibility and severity dimensions from the EPPM) interact with efficacy beliefs to predict behavioral responses, such that individuals experiencing high perceived health threat may engage more and apply different evaluation criteria to AI versus traditional search results. Hypothesis 5. Perceived threat will negatively predict behavioral intentions for AI-powered search while showing weaker or non-significant relationships with traditional search intentions.
Recent scholarship highlights opportunities to adapt TAM II for AI contexts where algorithmic transparency, perceived autonomy, and trust play increasingly significant roles.49,50 Usability remains key but new dimensions unique to AI—such as conversational interface quality and cognitive load—are increasingly important for future research.4,51 These additional constructs complement rather than replace TAM II’s predictive framework, representing contextual factors that may moderate relationships between core TAM II constructs and behavioral intentions.
Study Design and Hypotheses
While artificial intelligence (AI) is increasingly used for online health information, little is yet known about how people compare AI-powered tools to traditional search engines when seeking information about serious health issues like colorectal cancer. Previous studies have focused on general technology adoption, but few have examined how trust, perceived threat, and ease of use affect people’s willingness to use AI for health decisions. This gap matters because if patients do not trust or understand AI tools, they may make less informed choices about their health. This study addresses the gap by directly comparing user responses to Google search results and AI-generated responses for colorectal cancer symptoms, focusing on the roles of trust, threat, and usability.
The 2×2 factorial design compares technology types (traditional search engines format of Google vs. AI technology) across symptom concern levels (high vs. low concern scenarios), addressing critical gaps in understanding how health-information urgency moderates technology acceptance patterns. This design is particularly relevant for cancer prevention and early detection, as individuals’ technology preferences for symptom evaluation may influence screening behaviors and care-seeking timing, which are critical determinants of cancer outcomes. Participants viewed mock-up interfaces representing each search type—conversational AI chatbot responses versus traditional ranked search results—with health information content held constant across conditions to isolate interface effects. These varying levels of health urgency may fundamentally alter how individuals evaluate and trust different information sources, as threat perceptions can shift information-processing strategies and preferences for sources.
The manipulation uses search terms representing meaningful psychological distinctions: high concern scenarios included “family history colorectal cancer,” “blood in stool causes,” and “persistent abdominal discomfort,” while low concern scenarios focused on “occasional gassiness causes,” “minor digestive issues,” and “intermittent stomach discomfort.” This approach allows examination of technology preferences across different health urgency levels and how different situations activate evaluation criteria for AI versus traditional search methods. Hypothesis 6. The relationship between technology type and behavioral intentions will be moderated by concern level, with preferences for traditional search becoming more pronounced in high concern scenarios.
Figure 1 illustrates the psychological processes examined in H1-H5. H6, examining how concern level moderates technology type effects on behavioral intentions, was tested through factorial ANOVA rather than path analysis.
Methods
Participants
Data were collected by the research team as part of a larger statewide survey examining public opinion about a variety of health issues, including health communication patterns, healthcare access, and technology adoption behaviors. 52 Study procedures were approved by the Institutional Review Board (IRB) at The University of Texas at Austin (IRB Protocol #STUDY00002185, approved 23 December 2021); the study was also conducted in accordance with the Helsinki Declaration of 1875, as revised in 2024. This online survey was conducted in June–July 2024 and administered by the survey research firm Centiment. Centiment maintains research panels that include individuals who have agreed to be contacted for participation in online surveys and have been verified for survey quality and engagement. Participants offer written consent through Centiment in line with IRB best practices. All participant data were de-identified to protect privacy, and no individual identifying information is reported in any form that could enable identification of specific participants.
For the current study, Centiment recruited a sample of adult Texans (ages 18 and older) balanced on age, gender, and race/ethnicity to reflect the demographic composition of the state. Research panel members were screened for Texas residency, experience using internet search tools for information seeking, and basic familiarity with both traditional search engines and AI-driven interfaces, with quota sampling used to achieve demographic balance. No screening was conducted for prior colorectal cancer history, caregiver status, or health literacy levels, reflecting a general population sample. Eligible respondents were then randomly assigned to experimental conditions.
Upon study completion, participants received monetary compensation for their time directly from Centiment according to standard panel compensation rates. Data quality was ensured through multiple attention checks and screening for response patterns. After removing participants who failed attention checks (n = 42) or provided straight-lining answers across multiple construct measures (n = 38), the final analytical sample included 764 participants. No formal a priori power analysis was conducted to determine sample size. The sample size was determined by the broader survey project parameters.
Procedure
The prospective experimental study employed a 2 (Search Engine vs. Generative AI) × 2 (high concern vs. low concern scenario) between-subjects factorial design conducted through an online experimental platform. The reporting of this study conforms to the STROBE guidelines for transparent research reporting.
53
Randomization to conditions was conducted automatically through the survey platform to ensure equal distribution across the four experimental conditions. Figure 2 illustrates the four stimulus conditions and their content variations. Experimental stimuli for 2×2 factorial design
After providing informed consent and completing demographic questions, participants answered a set of general survey questions about their health information seeking behaviors, technology usage patterns, and healthcare experiences. Following these baseline measures, respondents were randomly assigned to one of the four experimental conditions where they viewed a stimulus image depicting either a traditional search engine results page or an AI interface response related to a colorectal cancer symptom scenario. Participants were instructed to imagine they were experiencing the symptoms described and to carefully examine the search results or AI response presented. After viewing their assigned stimulus, respondents completed the outcome measures related to their perceptions of the search method they had just observed. The survey concluded with additional questions about demographics and health characteristics not used in the current analysis.
Stimuli Creation
Stimuli development followed a systematic process to ensure realistic and representative examples of each search technology type. For the different stimuli images (Search Engine vs. Generative AI), specific search phrases were carefully selected to represent distinct levels of symptom concern. AI stimulus content was generated using GPT-4o (gpt-4o-2024-05-13) in May 2024, prior to the survey data collection period (June-July 2024). No specific model configurations or custom settings were applied beyond the standard Chat GPT interface available to general users. AI stimulus content was generated by entering the exact symptom-related prompts into ChatGPT (GPT-4) as simple search-style queries without additional instructions or contextual framing, mirroring how typical users might seek health information. High concern prompts included “family history colorectal cancer,” “blood in stool causes,” and “persistent abdominal discomfort.” Low concern prompts included, “occasional gassiness causes,” “minor digestive issues,” and “intermittent stomach discomfort.” These standardized prompts were used to generate the AI responses depicted in the stimulus screenshots.
Screenshots were captured from Google search results pages using standardized browser settings and screen resolution to ensure consistency. For Google search conditions, screenshots were captured using incognito browsing mode to prevent personalized results based on search history or user profiles/profile-based customization that might create researcher-specific bias. Consistent browser settings were maintained across all captures (Chrome browser, 1920x1080 resolution, standard zoom level (default font size), and all Google screenshots were captured during the same week in May 2024 to minimize temporal variation in search rankings and content availability. The Google screenshots included the typical search results layout with page titles, brief descriptions, source URLs, and related search suggestions. Participants viewed complete first-page Google results displaying multiple organic results in Google’s standard ranking format, typically including 7-10 results. No individual results were extracted or curated; participants viewed the authentic Google search interface as it would appear to a user entering these search terms during the stimulus development period.
For the AI condition, screenshots of ChatGPT responses were captured, showing the conversational interface and the AI’s detailed response to the symptom-related queries. All participants within AI conditions viewed identical ChatGPT-generated responses that provided synthesized information without source citations or references, representing standard ChatGPT interface output available during May 2024. Care was taken to select responses that were representative of typical AI output length and style, neither unusually brief nor excessively comprehensive. All screenshots were edited to remove any identifying information while maintaining the authentic appearance of each platform. Stimuli images were pilot tested with 25 individuals to ensure clarity and realism before inclusion in the main study. Special attention was paid to ensuring that the information quality and medical accuracy were comparable between conditions, with only the presentation format (traditional search results vs. conversational AI response) differing between experimental conditions. Figure 2 illustrates the four experimental conditions used in the 2×2 factorial design, showing representative examples of Google search results and AI conversational responses for both high concern and low concern colorectal cancer symptom scenarios.
Item Measures
This study used measures from the Technology Acceptance Model II 44 as updated by Kamal et al 54 to predict artificial intelligence usage. All measures used 7-point Likert scales ranging from 1 (strongly disagree) to 7 (strongly agree) unless otherwise noted. Additional measures were incorporated to capture constructs particularly relevant to health information seeking, including trust dimensions and threat perceptions that may be especially salient when individuals evaluate health-related technologies.
Perceived Usefulness
Perceived Usefulness was measured using four items adapted from Davis 44 and updated for the health information context: “Using this search method would improve my ability to find health information,” “Using this search method would enhance my effectiveness in understanding health issues,” “I would find this search method useful for health-related questions,” and “This search method would help me make better health decisions.” The original scale demonstrated high reliability (α = .87) in technology adoption contexts, and the health-adapted version showed similar reliability in pilot testing.
Perceived Ease of Use
Utilized four items based on the established TAM II scale: “Learning to use this search method would be easy for me,” “I would find it easy to get this search method to do what I want it to do,” “My interaction with this search method would be clear and understandable,” and “I would find this search method to be flexible to interact with.” These items have consistently demonstrated strong psychometric properties across a range of technological contexts (α = .91 in original validation studies).
Trust in Technology
Trust in Technology was assessed using a two-dimensional approach combining institutional trust and situational trust. Institutional trust (4 items) measured general confidence in the search technology provider: “I trust the organization behind this search method to provide accurate health information,” “This search method comes from a reliable source,” “I have confidence in the overall integrity of this search system,” and “I believe this search method operates with my best interests in mind.” Situational trust (4 items) focused on trust in specific search interactions: “I trust the health information provided by this search method,” “I would feel comfortable acting on health advice from this search method,” “I believe this search method provides unbiased health information,” and “I trust this search method to give me complete health information.”
Behavioral Intention
Behavioral Intention was measured through four items adapted for health information seeking contexts: “I intend to use this type of search method for health questions in the future,” “I predict I would use this search method for health information needs,” “I plan to use this search method when I have health concerns,” and “I would recommend this search method to others seeking health information.”
Attitude Toward Technology
Utilized five items capturing both cognitive and affective evaluations: “Using this search method for health information would be beneficial,” “Using this search method for health information would be wise,” “I like the idea of using this search method for health questions,” “Using this search method for health information would be pleasant,” and “Using this search method for health information would be satisfying.”
Perceived Threat
Perceived Threat was conceptualized using the Extended Parallel Process Model 13 framework, incorporating both susceptibility and severity components. Perceived Susceptibility (4 items) measured individuals’ assessment of their likelihood of experiencing negative consequences from using the search method: “I am at risk for receiving inaccurate health information from this search method,” “It is likely that I could be misled by this search method,” “I could easily get wrong health information from this search method,” and “There is a good chance I could make poor health decisions based on this search method.” Perceived Severity (4 items) assessed the perceived seriousness of potential negative outcomes: “Getting inaccurate health information would be serious,” “Making health decisions based on wrong information would be harmful,” “Being misled about health matters would have serious consequences,” and “Incorrect health information could significantly impact my wellbeing.”
Stress Related to Technology Use
Stress Related to Technology Use was measured using four items developed specifically for this study to capture anxiety and cognitive load associated with using different search methods: “Using this search method for health information would make me feel anxious,” “I would feel stressed trying to evaluate information from this search method,” “This search method would create uncertainty about health information quality,” and “I would worry about the reliability of information from this search method.”
All multi-item scales demonstrated acceptable to strong internal consistency: Perceived Usefulness (Google α = .884, AI α = .923), Perceived Ease of Use (Google α = .898, AI α = .901), Attitude (Google α = .936, AI α = .961), Behavioral Intention (Google α = .946, AI α = .951), Trust (Google α = .865, AI α = .861), Stress (Google α = .774, AI α = .848), and Threat (Google α = .827, AI α = .846). These reliability coefficients support the psychometric adequacy of the measures for hypothesis testing.”
Data Analysis
Data analysis proceeded through several phases to ensure appropriate handling of the experimental design and theoretical model testing. Initial analyses included frequency distributions and descriptive statistics for all variables to assess data quality and distributional properties. Exploratory factor analysis was conducted to confirm the factor structure of all multi-item scales, followed by reliability analysis using Cronbach’s alpha coefficients.
The primary analytical approach utilized two-way analysis of variance (ANOVA) to examine the main effects of search technology type (Google vs. AI) and symptom concern level (high vs. low), as well as their interaction effects, on key outcome variables including perceived usefulness, perceived ease of use, trust in technology, attitude toward technology, and behavioral intention to use technology. Effect sizes were calculated using eta-squared (η2) to assess practical significance of observed differences. Post-hoc analyses using Tukey’s HSD test were conducted when significant main effects were detected to identify specific group differences.
The factorial ANOVA design tested H6 (technology type × concern level interaction), while stepwise regression analyses examined the psychological pathways outlined in Figure 1 (H1-H5). Consistent with TAM II’s theoretical framework where perceived usefulness mediates relationships between external variables and behavioral intentions, stepwise multiple regression analyses were conducted separately for each technology condition (Google search and AI search) to identify the most important predictors of behavioral intention within each technological context. Predictor variables included all TAM II constructs (perceived usefulness, perceived ease of use, attitude), trust measures, threat perceptions (susceptibility and severity), and stress related to technology use. The stepwise approach allowed for identification of the most parsimonious models while controlling for multicollinearity among predictors. Regression assumptions were tested through examination of residual plots, normality tests, and variance inflation factors (VIF < 3.0 for all predictors).
Results
Participant Demographics
Demographics
Educational attainment varied across participants, with 28.7% having completed high school or equivalent, 31.4% having some college education, 25.3% holding bachelor’s degrees, and 14.6% having completed graduate or professional degrees. Annual household income distribution showed 22.8% earning less than $30,000, 28.1% earning $30,000-$59,999, 26.4% earning $60,000-$99,999, and 22.7% earning $100,000 or more. Healthcare access patterns indicated that 78.4% of participants had regular healthcare providers, while 21.6% reported lacking consistent access to medical care.
Descriptive Statistics for Key Variables
Descriptive Statistics for Study Variables (N = 764)
Note. All variables measured on 7-point Likert scales.
Perceived usefulness of the search technologies received positive evaluations (M = 4.64, SD = 1.40), with participants generally viewing both AI and traditional search as beneficial for health information seeking. Attitudes toward the search technologies were similarly positive (M = 4.61, SD = 1.54), while behavioral intentions to use the technologies showed moderate to strong endorsement (M = 4.56, SD = 1.69).
Stress related to using search technologies for health information was moderate (M = 3.89, SD = 1.41), suggesting that participants experienced some anxiety about technology-mediated health information seeking but that this concern was not overwhelming. Perceived threat associated with search technology use was moderate to high (M = 4.67, SD = 1.18), indicating meaningful concern about potential risks associated with online health information seeking.
The threat components showed noteworthy patterns: perceived severity of negative outcomes from inaccurate health information was high (M = 5.61, SD = 1.34), reflecting participants’ recognition that health information quality has serious implications. However, perceived susceptibility to receiving inaccurate information was more moderate (M = 3.72, SD = 1.54), suggesting that while participants understood the potential consequences of misinformation, they felt somewhat confident in their ability to avoid it.
Experimental Condition Effects
A 2×2 factorial ANOVA examined the effects of search interface type (Google vs. AI) and symptom concern level (low vs. high concern) on technology acceptance measures. For trust in health information search technology, there was a significant main effect for search type, F(1,758) = 7.26, p = .007, η2 = .009, with Google search receiving higher trust ratings (M = 4.15, SD = 1.06) compared to AI search (M = 3.93, SD = 1.16). No significant main effect emerged for concern level, F(1,758) = 0.38, p = .537, and no interaction occurred between search type and concern level, F(1,758) = 0.01, p = .919. Given the absence of significant concern level main effects across all dependent variables, including stress, threat perceptions, behavioral intentions, and other TAM II constructs (all p > .24), results are reported collapsed across concern conditions in subsequent analyses. These results indicate that established familiarity with Google’s search approach may contribute to higher trust perceptions, or that participants harbor specific concerns about AI-generated health information.
Technology Type Comparisons
Means, Standard Deviations, t-tests, and Effect Sizes Comparing Google Search and AI Search Conditions
Perceived usefulness also favored Google search over AI search (t = 3.89, p < .001), indicating that participants viewed traditional search engines as more beneficial for their health information needs. This difference may reflect perceptions about information comprehensiveness, source transparency, or confidence in evaluation criteria for traditional versus AI search. Attitudes toward the search technologies showed significant preference for Google over AI (t = 5.98, p < .001), with participants expressing more positive overall evaluations of traditional search approaches. Behavioral intentions followed a similar pattern (t = 9.38, p < .001), with participants indicating significantly stronger intentions to use Google search compared to AI for future health information seeking.
No significant differences emerged between conditions for stress (t = -0.32, p = .748), overall threat perceptions (t = 1.43, p = .151), perceived severity (t = 1.45, p = .147), or perceived susceptibility (t = 0.95, p = .340). These non-significant findings suggest that while participants differentiated between the two technologies on functional dimensions (ease of use, usefulness) and evaluative dimensions (trust, attitude, behavioral intention), they perceived similar levels of risk and emotional concern associated with both approaches.
Correlational Analyses
Correlations Among Study Variables: Google Search Condition (n = 382)
Correlations Among Study Variables: AI Search Condition (n = 382)
Similar patterns emerged for the AI search condition. Perceived usefulness of AI correlated strongly with intention to use AI (r = .79, p < .001), as did ease of use (r = .72, p < .001). However, there was a stronger relationship between trust and ease of use in the AI condition (r = .60, p < .001) as compared to the Google search condition (r = .39, p < .001). The relationship between trust and intention to use AI search (r =.72, p<.001) was stronger than the relationship between trust and intention to use Google search (r =.58, p<.001).
Stress related to technology usage had a significant negative relationship with trust (r = -.24, p <.001), ease of use (r = -.15, p<.001), usefulness (r = -.14, p<.001), and intention to use AI (r= -.25, p<.001). However, none of these relationships were statistically significant in the Google search condition. These findings indicate that stress was negatively correlated with trust, ease of use, usefulness, and intention to use AI, but not with Google search, suggesting that participants experienced greater uncertainty and reliability concerns when using AI search compared to Google search.
Regression Analyses
Stepwise regression analyses were conducted to identify the most important predictors of perceived usefulness across both technology conditions. This approach was used within a theory-driven framework to identify the most parsimonious TAM II models while accounting for multicollinearity among conceptually related constructs. Variable entry was limited to established TAM II constructs and theoretically motivated extensions rather than exploratory, data-driven variable selection. The initial regression model contained only perceived severity as a predictor (β = .28, p < .001) and explained 7.6% of the variance in perceived usefulness (R2 = .08, F(1,758) =62.08, p < .001). In the final model (R2 = .080, ΔR2 = .004), severity remained the strongest predictor (β = .25, p < .001) of perceived usefulness, and perceived susceptibility added a smaller but significant contribution (β = .07, p = .048). Analyses examined theoretical predictors without demographic controls, focusing on TAM II threat perception variables. Figure 3 illustrates the regression pathways among ease of use, usefulness, and search type predicting attitude toward technology. Regression paths among ease of use, usefulness, and search type predicting attitude toward technology
Mediation Analyses
Mediation analyses were conducted separately before moderated mediation to first establish basic indirect effects, then test whether these pathways differed by technology type, following recommended practices for sequential mediation analysis. 55 Hayes PROCESS Model 4 examined trust, attitude toward technology, and intention to use technology. For both Google search and AI search, there was a significant mediation effect. Furthermore, it is important to note that trust plays a much greater role in the AI condition than the google search condition.
Greater trust in Google search (X) leads to greater attitude toward Google search (M) (B=.84, SE= .05, t=16.53, CI from .74 to .94), and greater attitude toward Google search leads to greater intention to use Google search (Y) (B=.71, SE= .05, t=14.6, CI from .61 to .80). There was a significant direct effect indicating that greater trust leads to greater intention to use Google search (B=0.17, SE=.06, t=2.63, CI from 0.04 to 0.29). The results indicated that attitude toward Google search (M) partially mediates the relationship between trust (X) and intention to use Google search (Y) (Figure 4). Google condition only regression paths among trust, attitude toward technology, and behavioral intention
Greater trust in AI search (X) leads to greater attitude toward AI search (M) (B=.99, SE= .05, t=19.47, CI from .89 to 1.09), and greater attitude toward AI search leads to greater intention to use AI search (Y) (B=.73, SE= .04, t=19.22, CI from .66 to .81). There was a significant direct effect indicating that greater trust leads to greater intention to use AI search (B=0.33, SE=.05, t=6.2, CI from 0.23 to 0.43). The results indicated that attitude toward AI search (M) partially mediates the relationship between trust (X) and intention to use AI search (Y) (Figure 5). AI condition only regression paths among trust, attitude toward technology, and behavioral intention
Summary of Direct, Indirect, and Total Effects From Mediation and Moderated Mediation Analyses
Abbreviations: B, unstandardized regression coefficient; CI, confidence interval; SE, standard error.
a Bootstrapped indirect effects based on 5000 bootstrap samples. Confidence intervals that exclude zero indicate a statistically significant indirect effect. For moderated mediation, Search type is the moderator (0 = Google Search, 1 = AI Search). Conditional effects represent estimated effects at each level of the moderator.
Moderated Mediation Analysis
Moderated mediation analysis was conducted to test the central theoretical proposition that technology type would moderate the indirect effect of perceived threat on behavioral intention through perceived usefulness, consistent with extended TAM II frameworks examining contextual moderators of technology acceptance pathways. Hayes PROCESS Model 15 was used with ease of use, usefulness, attitude toward technology, and search type (Hayes, 2022).
Perceived ease of use (X) predicted higher perceived usefulness (M) (B=.63, SE= .03, t=24.30, CI from .58 to .68).Perceived usefulness predicted higher attitude toward the technology (Y) (B=.70, SE= .10, t=7.14, CI from .51 to .90). There was no significant direct effect of perceived ease of use on attitude toward technology (B=.06, SE=.10, t=0.62, CI from -0.13 to 0.25). The results indicated that perceived usefulness (M) completely mediates the relationship between perceived ease of use (X) and attitude toward technology (Y).
The conditional direct effect of perceived ease of use on attitude differed by search type: Google search (B=0.25, SE=.04, CI from 0.16 to .33) and AI search (B=0.43, SE=.04, CI from .36 to .51). Figure 6 displays the interaction effect of attitude and ease of use by search type. The index of moderated mediation did not differ significantly from zero (index=-.04, SE=.11, CI from -.15 to .05), indicating no evidence that the indirect effect via usefulness varied by search type. Taken together, the pattern is consistent with partial mediation with a search-type–moderated direct path. Interaction effect of attitude and ease of use by search type
Scenario Intensity Effects
Two-way ANOVAs tested whether symptom concern level (high vs. low) moderated the relationships between search type and outcomes (ease of use, usefulness, attitude, intention, trust). No significant two-way interactions were observed for any dependent variable.
Hypothesis Testing Summary
The study’s hypotheses received mixed support. H1 and H2 were supported, with perceived usefulness (r = .79, p < .001) and ease of use (r = .72, p < .001) showing strong positive relationships with AI adoption intentions. H3 regarding trust moderation was supported through mediation analyses showing trust’s central role in technology acceptance. H4 was supported, as trust emerged as a stronger predictor of behavioral intention for AI search (r = .74) compared to Google search (r = .61). H5 was partially supported, as perceived threat showed significant negative correlations with AI acceptance measures but not with Google search acceptance. H6 was not supported, with no significant interaction between technology type and concern level (F = 0.01, p = .92), indicating consistent technology preferences regardless of symptom severity.
Discussion
This study reveals substantial differences in how individuals evaluate AI versus traditional search engines for colorectal cancer information seeking, with critical implications for cancer control efforts and patient care outcomes. These differences extend beyond simple technology preferences to fundamental questions about how patients access, evaluate, and act upon cancer-related health information in digital contexts, potentially influencing screening behaviors, symptom interpretation, and care-seeking decisions.
The most striking finding was the overall preference for Google search over AI across all key dimensions measured by the Technology Acceptance Model II, despite AI’s theoretical advantages in providing conversational responses. For colorectal cancer information seeking specifically, this preference pattern suggests that patients may be more likely to trust and act upon traditional search results when evaluating gastrointestinal symptoms or screening recommendations, with potential consequences for early detection and preventive care.
These technology preferences have direct implications for cancer control, as individuals’ choices between AI and traditional search when evaluating symptoms like changes in bowel habits or abdominal discomfort may influence screening uptake, care-seeking timing, and symptom interpretation. The preference for traditional search suggests that cancer prevention campaigns utilizing digital channels should consider these technology acceptance patterns when designing information delivery strategies.
The substantial difference in perceived ease of use (t = 8.02, p < .001) indicates that participants found traditional search interfaces more intuitive than AI conversational systems. This finding challenges assumptions about AI’s user-friendliness and indicates that conversational interfaces may introduce cognitive complexity that outweighs their apparent naturalness (Bubaš et al, 2024). The ease-of-use advantage for Google search likely reflects years of learned interaction patterns but may also suggest fundamental usability barriers in current AI health information interfaces.
Trust emerged as a more critical factor for AI adoption than for traditional search engine use, with trust showing stronger correlations with behavioral intention for AI (r = .72) compared to Google search (r = .55). This pattern suggests that AI systems face a higher threshold for acceptance in health contexts, where users may apply more stringent evaluation criteria or users may default to known search tools because of personal histories. Additionally, as a newer, more polarizing form of technology, AI tools have the ability to invoke stronger responses regarding trust, usage, and acceptance.
Trust and Authority in AI Health Information
The differential role of trust across technology conditions reveals important insights about how individuals conceptualize AI versus traditional search in health contexts. These findings align with broader work investigating the establishment of trustworthy AI in healthcare contexts. 56 Traditional search engines benefit from what might be termed “procedural trust,” thought of as confidence in the search process system and users’ ability to evaluate multiple sources. AI systems, conversely, require “epistemic trust,” a confidence in the system’s knowledge and reasoning capabilities as an individual source.
This distinction has profound implications for AI adoption in healthcare settings. While traditional search allows users to maintain agency in information evaluation and synthesis, AI systems ask users to delegate aspects of this cognitive work to algorithms. The higher trust threshold for AI adoption likely reflects users’ awareness of this delegation and their uncertainty about AI systems’ medical knowledge boundaries.
The finding that participants perceived similar levels of threat and stress across both conditions, despite clear preferences for Google search, suggests that concerns about AI in health contexts may be more cognitive than emotional. Participants appear to recognize AI’s potential utility while maintaining reservations about its reliability and appropriateness for health decisions.
Technology Acceptance Model II in the AI Era
The TAM II framework demonstrated robust explanatory power across both technology conditions. However, the predictive patterns differed meaningfully between conditions, suggesting that AI adoption activates distinct psychological processes compared to traditional technology acceptance.
The prominence of trust as the strongest predictor for AI intentions compared to its secondary role for Google intentions indicates that traditional TAM II applications may underestimate trust’s importance in contexts involving artificial intelligence. This finding suggests that TAM II may require modification when applied to AI technologies, particularly in high-stakes domains like healthcare. AI acceptance models may need to incorporate affective responses more explicitly, recognizing that AI technologies may trigger different emotional reactions than older digital tools. Further, as our trust measures captured both technology dependability and result trustworthiness, future research could benefit from examining specific trust dimensions (competence, integrity, benevolence, transparency) to provide more targeted guidance for AI system design and user acceptance interventions.
Clinical and Public Health Implications
These findings have immediate relevance for healthcare providers who may increasingly encounter patients influenced by health information from AI searches. The study suggests that patients using AI for health information seeking may arrive at clinical encounters with different knowledge structures and confidence levels compared to those using traditional search methods. Specifically, patients using AI search may possess more synthesized but less source-diverse knowledge, having received algorithmically-generated summaries rather than evaluating multiple competing sources, potentially affecting their ability to articulate information uncertainty or identify knowledge gaps.
These technology preferences also have implications for cancer control, as choices between AI and traditional search when evaluating symptoms may influence screening uptake, care-seeking timing, and symptom interpretation. The preference for traditional search may suggest that cancer prevention campaigns utilizing digital channels should consider these technology acceptance patterns when designing information delivery strategies. While current findings reveal acceptance challenges, AI technology also holds substantial promise for addressing known limitations of traditional search, including information overload, complex source evaluation demands, and the health literacy requirements for synthesizing multiple sources. Future AI implementations that address transparency and usability concerns while preserving conversational advantages may achieve greater acceptance.
Provider Training and Continuing Education Needs
Cancer care providers and public health professionals will require specialized training and continuing education to navigate the evolving landscape of AI-mediated health information seeking. Oncologists, primary care physicians involved in cancer screening, and cancer prevention specialists should be prepared to address AI-specific concerns such as algorithmic transparency, source verification, and the limitations of AI-generated medical guidance. They also will need to recognize that patients may have varying levels of trust in AI-generated information.
Public health practitioners in cancer control programs will similarly need enhanced competencies in: 1) evaluating AI-generated health communications, 2) understanding how these technologies shape community health information environments, and 3) developing cancer screening and prevention interventions that account for AI’s influence on cancer-related health decision-making. Given our findings showing greater trust in traditional search engines, providers should be prepared to discuss both AI-generated and traditional search results when counseling patients about colorectal cancer symptoms, screening recommendations, and risk factors.
Implications for Technology Development and Health Promotion
The preference for traditional search engines, despite AI’s conversational advantages, indicates that current AI implementations may not optimally serve individuals’ health information needs. This gap represents both a challenge and an opportunity for AI developers working in healthcare contexts such as cancer care. The findings suggest that successful health AI systems will need to address usability concerns while building trust through transparency and source attribution.
From a health promotion perspective, the results highlight the need for digital health literacy initiatives that specifically address AI technologies. Traditional digital health literacy frameworks focused on evaluating website credibility and source authority may be insufficient for navigating health information from AI searches, which requires different evaluation strategies such as evaluating algorithmic transparency, recognizing AI limitations, and understanding when to seek human expert verification.
Technology Preferences in Health Information Contexts
The differential trust and acceptance patterns observed between Google and AI search have particular implications when considered in the context of specific health conditions. The study’s focus on colorectal cancer information seeking provides insight into how technology preferences may vary with health condition characteristics. Colorectal cancer symptoms often involve sensitive bodily functions that individuals may prefer to research privately before consulting healthcare providers. The preference for Google search in this context may reflect users’ desire to maintain control over information sources and evaluation processes when dealing with potentially stigmatizing or anxiety-provoking symptoms. Traditional search results allow individuals to manage uncertainty gradually by choosing which sources to examine, when to stop reading, or how deeply to explore concerning information, 57 whereas AI’s synthesized responses may present potentially overwhelming information in a single, less controllable format. 11 The observed preferences may reflect users’ familiarity with link-based information evaluation paradigms rather than resistance to AI technology itself, suggesting that acceptance patterns stem from interaction model differences rather than technological novelty alone.
This preference for familiar search methods is further explained by differences in information transparency. Traditional search engines provide explicit source attribution, allowing users to assess information credibility based on institutional affiliations, author expertise, and publication context. AI systems, conversely, synthesize information from multiple sources without transparent attribution, potentially creating uncertainty about information provenance. This opacity may be particularly problematic for health information since source credibility strongly influences user trust and behavioral responses, as evidenced by our participants’ significantly higher trust ratings for Google search compared to AI search. This aligns with our finding that trust was the strongest predictor of behavioral intention (r = .72), suggesting that AI systems’ lack of source transparency may be a fundamental barrier to adoption.
These trust findings align with documented concerns among cancer patients about AI in clinical contexts, including fears of misdiagnosis, privacy breaches, and clinical accuracy. 58 Cancer patients’ documented discomfort with AI diagnostic applications corresponds with our observation that trust barriers persist even for information-seeking contexts, suggesting that AI acceptance challenges may be particularly pronounced in cancer care settings where accuracy concerns are heightened.
The absence of significant interaction effects between technology type and symptom concern level (high vs. low concern scenarios) suggests that technology preferences are more fundamental than context-specific. Participants consistently preferred Google over AI regardless of whether they were presented with mild or serious symptom scenarios. This finding indicates that interventions to improve AI acceptance in health contexts should focus on core usability and trust factors such as improving perceived ease of use, demonstrating clear usefulness benefits, and enhancing transparency to build user trust rather than tailoring approaches to specific symptom types or severity levels.
Limitations and Future Research Directions
Several methodological limitations constrain the interpretation of these findings. The use of static screenshots, while ensuring experimental control, cannot capture the interactive, iterative nature of AI conversations that may be central to their value proposition. Additionally, while incognito mode reduced algorithmic personalization, complete elimination of search result variation presents ongoing methodological challenge. However, because the study focused on interface paradigm differences rather than specific content variations, this limitation is unlikely to undermine the validity of the observed preference patterns. Future research should employ more ecologically valid designs that allow participants to engage in actual conversations with AI systems and compare these experiences to genuine search sessions, which could also mean opportunities for interdisciplinary collaborations with computer-science colleagues responsible for technological development.
The study’s cross-sectional design precludes causal inferences about technology adoption patterns, and the focus on intentions rather than actual behavior may limit conclusions about real-world adoption. Further, because a formal a priori power analysis was not conducted, smaller effects or subtle interaction patterns may have gone undetected. Longitudinal research tracking individuals’ evolving relationships with AI health tools would provide valuable insights into how initial perceptions develop over time and with experience. Our findings suggest that source transparency and cognitive demands may explain the preference for traditional search, but these mechanisms were inferred from the pattern of results rather than directly measured over time. Future research should explicitly assess perceptions of source transparency and cognitive load to test these proposed explanations for AI acceptance patterns.
Additionally, the study did not systematically assess several potentially important moderating variables, including prior AI familiarity, digital health literacy, health anxiety levels, educational background effects, and previous cancer history, which may also shape technology acceptance in ways this analysis could not capture. Future research should examine these factors alongside longitudinal changes to provide a more precise understanding of the mechanisms driving technology preferences in health contexts.
Further, the study employed hypothetical colorectal cancer symptom scenarios rather than participants’ actual health concerns. While this approach ensured experimental control, it may not fully capture how individuals would evaluate AI versus traditional search when facing genuine health anxieties or personal medical decisions. Future research should examine technology preferences in naturalistic settings where participants are seeking information about their own health symptoms.
The geographic restriction to Texas residents, while providing demographic representativeness within the state, may limit generalizability to regions with different healthcare systems, technology adoption patterns, or cultural attitudes toward AI. Additionally, the general population sample limits generalizability to actual colorectal cancer patients, who may demonstrate different technology acceptance patterns based on personal health experiences and disease-specific health literacy. Replication across geographic and cultural contexts would strengthen confidence in the findings.
Future research could also examine whether the observed patterns persist as AI technologies become more familiar and sophisticated. The study was conducted during the early phase of consumer AI adoption, when participants likely had extensive Google experience but limited AI exposure, potentially influencing preference patterns. Digital literacy levels and prior AI familiarly were not systematically measured in the current study yet may represent important variables to explore in future research especially as these technologies become more mainstream.
Research designs that allow manipulation of AI system characteristics—such as transparency about sources, confidence indicators, or explicit uncertainty acknowledgments—could identify specific features that enhance trust and adoption. Such studies would provide actionable guidance for AI developers working in healthcare contexts. This is particularly important for cancer-related information seeking, where AI systems must appropriately convey uncertainty about symptom significance and emphasize the need for professional medical evaluation when discussing potential warning signs.
Practical Recommendations
For AI System Developers
AI health information systems should prioritize transparency features that address the source attribution advantages of traditional search engines. This could include citations, confidence indicators, and clear acknowledgment of limitations. User interface design should minimize cognitive load while maintaining the conversational advantages that differentiate AI from traditional search.
Based on these findings, we recommend specific improvements to AI health information systems: (1) implementing source attribution and confidence indicators to enhance trust; (2) developing hybrid interfaces that preserve user control while offering conversational benefits; (3) providing transparent uncertainty communication and clear medical advice boundaries; and (4) incorporating user education components to support effective AI interaction skills.
For Healthcare Providers
Clinical training programs should incorporate modules on AI-generated health information, helping providers recognize when patients may have consulted AI systems and develop strategies for addressing AI-specific concerns such as potential misinterpretation of symptom severity or delay in seeking appropriate screening. Providers should be prepared to discuss the strengths and limitations of different information sources without dismissing patients’ technology preferences.
For Health Communicators and Public Health Agencies
Digital health literacy initiatives should expand beyond traditional website evaluation criteria to include AI-specific competencies. Public health agencies and professionals focused on cancer prevention and control should consider developing guidelines for effective AI health information seeking while maintaining appropriate boundaries around medical advice, particularly for cancer screening recommendations where AI systems should direct users to evidence-based guidelines and healthcare providers rather than providing screening advice.
For Healthcare Systems
Healthcare organizations should develop policies regarding AI health information tools, potentially including institutional AI implementations that provide the conversational benefits of AI while maintaining appropriate clinical oversight and source transparency.
Conclusion
The integration of AI into health information seeking represents a fundamental shift requiring both technological adaptation and reconceptualization of digital health literacy frameworks, especially for cancer prevention and screening since accurate symptom interpretation and timely care-seeking are critical for patient outcomes. While current AI implementations face substantial barriers related to trust, usability, and transparency, the technology still holds significant potential for personalized, accessible health communication.
Realizing this potential will require addressing users’ legitimate concerns about the reliability of AI output while preserving the agency and source transparency that characterize effective health information seeking. Our findings suggest that AI adoption in healthcare may not simply replicate patterns observed in other domains; instead, it necessitates specific attention to the unique characteristics of health information needs, including privacy concerns, risk sensitivity, and the critical importance of accuracy.
As AI technologies continue to evolve, ongoing research will be essential to understand optimal integration into the complex ecosystem of health information seeking and healthcare delivery. Future research should prioritize real-world evaluations of AI health information tools, focusing on interventions that enhance trust, usability, and transparency. Collaborative efforts between technologists, clinicians, and behavioral scientists may be an important strategy to ensure that AI systems meet the needs of varied patient populations.
Footnotes
Acknowledgements
The authors thank the participants who contributed their time to this research. AI assistance was used for language editing and manuscript formatting, but all research content, analysis, and conclusions remain those of the authors.
Ethical Considerations
This study was approved by the Institutional Review Board at The University of Texas at Austin (IRB Protocol # STUDY00002185, approved 23 Dec 2021).
Consent to Participate
Informed consent was obtained from all participants through the online survey platform prior to data collection.
Author Contributions
Brad Love: Conceptualization, Methodology, Formal Analysis, Validation, Investigation, Writing: Original Draft, Writing: Review & Editing, Visualization, Project Administration, Supervision. Charulata Ghosh: Conceptualization, Methodology, Formal Analysis, Software, Investigation, Data Curation, Writing: Original Draft, Writing: Review & Editing, Visualization, Project Administration. Weijia Shi: Conceptualization, Methodology, Formal Analysis, Software, Investigation, Data Curation, Writing: Original Draft, Writing: Review & Editing, Project Administration. Karly Quaack: Conceptualization, Validation, Writing: Original Draft, Writing: Review & Editing, Visualization, Project Administration. Michael Mackert: Conceptualization, Investigation, Data Curation, Writing: Original Draft, Project Administration, Funding Acquisition, Supervision.
Funding
The authors received no financial support for the research, authorship, and/or publication of this article.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Data Availability Statement
Data are available upon request from the corresponding author.
Use of Artificial Intelligence Statement
During the preparation of this work the author(s) used Claude Sonnett 4.5 (accessed April 2026) in an assistive manner to check grammar, spelling, and language in line with Sage’s guidelines <
>. After using this tool, the authors further reviewed and edited the content as needed and take full responsibility for the content of the publication.