AI, particularly generative AI (GenAI), has rapidly transformed the health care landscape, offering new tools for patient education, clinical decision support, and personalized medical recommendations. Since the release of ChatGPT in late 2022, public engagement with GenAI systems has surged, with a growing number of individuals turning to AI-powered chatbots for health-related queries [1,2]. These tools promise convenience and accessibility, but they also produce erroneous outputs, raising critical questions about how users evaluate and calibrate trust in AI-generated health information [3-14]. Studies demonstrate that AI chatbot responses to health queries frequently deviate from clinical guidelines, contain inaccuracies, and may mislead patients who are unaware of these limitations [15,16]. A key concern emerging from this pattern of use is that repeated and largely uncritical engagement with GenAI may foster overreliance. As users experience GenAI outputs as convenient and immediately accessible, they may progressively reduce their own evaluative effort, a process that, over time, can give rise to learned dependency. Learned dependency is a behavioral phenomenon defined as habitual reliance on external AI systems arising from repeated reinforcement, often at the expense of independent thinking and critical evaluation [17]. Research on overreliance on AI demonstrates that users frequently accept incorrect AI outputs, particularly when they perceive the system as competent [18-22]. This pattern, known as automation bias, is strengthened through repeated reliance, as habitual use of AI systems progressively reduces users’ critical scrutiny of system outputs [18,22]. In health information contexts, where the consequences of misplaced trust may directly affect health decisions, understanding how learned dependency shapes trust calibration is a pressing but underexplored concern.

Despite the rapid integration of GenAI into health information ecosystems, existing research has not adequately examined how learned dependency shapes trust calibration in this context. Prior work has explored dependency-related phenomena among students [23] and professionals such as clinicians or software developers [24], yet health information consumers whose GenAI use directly bears on health comprehension and decision-making have remained largely overlooked. Prior research on online health information has extensively documented how source credibility, interface design, and information quality shape user trust [25], yet the role of habitual AI dependency in disrupting trust calibration has not been experimentally examined. Critically, no prior experimental study has tested whether learned dependency disrupts users’ ability to distinguish accurate from inaccurate AI-generated health information, or whether interface-level interventions such as text-based visual attention cues (TVCs) can counteract such effects. Addressing this gap is essential for designing AI-mediated health systems that enhance rather than erode users’ critical evaluation [26].

Our study addresses this gap through 2 controlled experiments examining how learned dependency on GenAI affects trust calibration in AI-generated health information. We focus on trust as the key outcome because it is the primary cognitive mechanism through which users decide whether to accept or question AI-generated content, and because prior research consistently identifies trust as central to health information engagement and AI adoption [3,27-30]. We further test whether text highlighting, a practical interface-level attention cue, can mitigate dependency-driven overreliance on inaccurate AI outputs. By focusing specifically on trust as the outcome variable, this study makes a deliberate and bounded contribution: we do not examine downstream behavioral intentions or adoption behavior, which represent important but distinct constructs that warrant dedicated investigation in future work.

Our study draws on 2 complementary theoretical frameworks: trust calibration theory and attention theory. Trust calibration theory holds that effective human-automation interaction requires users to adjust their trust in proportion to a system’s actual reliability [27]. Well-calibrated trust enables users to rely on a system when it performs correctly and withhold trust when it errs. Critically, miscalibrated trust in either direction has consequences: overreliance leads users to accept erroneous outputs, while underreliance causes them to discard useful information [27,28]. This framework underpins our prediction about the connection between information accuracy and trust.

Building on these theoretical foundations, we introduce learned dependency as a behavioral phenomenon that may disrupt trust calibration. Learned dependency refers to the tendency for users to increasingly rely on AI-generated outputs after repeated exposure and perceived successful use. This construct is theoretically grounded in 2 related literatures. First, research on automation bias demonstrates that repeated reliance on automated systems leads users to defer to system outputs with reduced critical scrutiny, even when errors are present [18,20,21]. This effect is particularly pronounced in health contexts, where AI-generated advice has been shown to significantly reduce diagnostic accuracy when it is incorrect [21]. Second, cognitive offloading theory suggests that users who habitually delegate information evaluation to external systems progressively reduce their own cognitive investment in that task [22,31], a pattern that becomes self-reinforcing over repeated interactions. Together, these mechanisms suggest that learned dependency may be positively associated with overall trust in AI outputs while simultaneously weakening users’ sensitivity to information inaccuracy, as dependent users process AI outputs with reduced analytical engagement [32].

Against this backdrop, interface-level attention cues such as text highlighting represent a practical and scalable intervention to support trust calibration. Highlighting key information increases its perceptual salience, which, according to the limited capacity model of mediated message processing [33], should free attentional resources for evaluative processing rather than information search. Eye-tracking research has shown that users consistently direct their visual attention toward highlighted text areas, demonstrating the von Restorff isolation effect for digital highlighting interfaces [34,35]. The von Restorff effect, also known as the isolation effect, suggests that when people are presented with several similar items, the item that stands out as different is more likely to capture attention and be remembered. In digital highlighting interfaces, this principle works by making highlighted information visually distinct, prompting users to pause and process it more carefully. Importantly, among available highlighting methods, bold text has been found to produce the shortest visual search times in graphical user interface tasks, outperforming color-only highlighting on search efficiency [36]. Whether such attentional cueing is sufficient to support trust calibration among highly dependent users, however, remains an open empirical question. We pose our research questions as follows:

Research question 1: How do the accuracy of AI-generated health information and TVC (eg, highlighting text information) influence health information consumers’ trust in AI-generated information?

Research question 2: How does learned dependency on GenAI influence health information consumers’ trust in AI-generated information, and how does it moderate the effects of information accuracy and TVC on trust?

Building on the integrative framework presented above, we examine how information accuracy, TVC, and learned dependency on GenAI jointly shape health information consumers’ trust in AI-generated content. Rooted in trust calibration theory and attention theory, our framework proposes that habitual reliance on GenAI reduces scrutiny in information evaluation, with important consequences for how users respond to accurate and inaccurate AI-generated outputs. We have developed the research framework (Figure 1), and the corresponding hypotheses are presented below.

Trust calibration theory predicts that users should, under normal conditions, trust accurate information more than inaccurate information. When AI-generated health information is correct, users who evaluate it critically are expected to recognize its quality and report higher trust. We therefore hypothesize the following:

H1: Information accuracy is positively associated with health information consumers’ trust in AI-generated health information.

Attention theory suggests that directing users’ attention to specific portions of AI-generated health information through TVC can support more careful evaluation of that content. In this study, text highlighting is applied to key portions of the AI-generated health text regardless of whether those portions contain accurate or inaccurate information, depending on the experimental condition. By directing attentional resources toward the highlighted content, users are better positioned to compare it against the source clinical text and detect potential inaccuracies. Text highlighting is, therefore, proposed as a design intervention that supports trust calibration rather than simply increasing or decreasing trust; that is, users who receive highlighting should be more sensitive to information accuracy than those who do not. Therefore, we propose the following hypothesis:

H2: The presence of TVC (text highlighting) is positively associated with health information consumers’ trust in AI-generated health information.

Research on automation bias and cognitive offloading suggests that users who have developed habitual reliance on AI systems tend to engage with AI outputs with reduced analytical effort. As a result, highly dependent users are expected to extend greater overall trust in AI-generated information, regardless of its accuracy. This leads to the following hypothesis:

H3: Learned dependency on GenAI is positively associated with health information consumers’ trust in AI-generated health information.

Although H3 predicts a positive main effect of learned dependency on trust, trust calibration theory further predicts that this dependency may distort the relationship between information accuracy and trust. Specifically, users with high learned dependency are expected to process AI outputs with reduced scrutiny, weakening their ability to distinguish between accurate and inaccurate content. Based on this reasoning, we hypothesize the following:

H4: Higher learned dependency is negatively associated with the positive relationship between information accuracy and health information consumers’ trust in AI-generated health information, such that the effect of information accuracy on trust is weaker when learned dependency is higher.

Finally, although TVCs may generally support trust calibration (H2), their effectiveness may be attenuated among users with high learned dependency. Highly dependent users may continue to accept AI outputs with limited scrutiny even when key information is visually highlighted, as habitual reliance reduces their motivation to engage in effortful evaluation. We therefore hypothesize:

H5: TVCs weaken the trust-distorting effect of learned dependency on trust calibration such that the negative moderating effect of learned dependency on the relationship between information accuracy and trust (H4) is attenuated when TVCs are present.

Figure 1 illustrates the research framework guiding this study. Information accuracy and TVC are positioned as AI-generated information characteristics that directly influence trust. Learned dependency is positioned as an individual trait that both directly influences trust (H3) and moderates the effects of information accuracy (H4) and TVC (H5) on trust. The framework reflects a 2 × 2 experimental design in which information accuracy and TVC are manipulated between participants, while learned dependency is measured as a continuous individual difference variable.

Our study contributes to the literature in several ways. First, it advances research on human-AI interaction by introducing learned dependency as a behavioral phenomenon that can influence trust calibration in AI-generated health information. While prior work has examined trust in AI systems [19,37,38], automation bias [18,20], cognitive forcing functions [19], and GenAI dependency [26] separately, no prior study has experimentally tested their joint influence on trust calibration in health information contexts.

Second, the study tests whether text highlighting is a practical interface-level attention cue that can attenuate the trust-distorting effect of learned dependency. Building on prior work on cognitive forcing functions [19] and typographical cues, specifically TVC [35,36,39-41], the study extends this literature to GenAI health information contexts and provides evidence on the limits of surface-level attentional interventions.

Third, the study offers practical implications for the design of GenAI systems used in health information environments. As consumers increasingly rely on GenAI tools for medical guidance and interpretation of health information, understanding how trust and dependency shape user behavior becomes critical to the safety of health information consumers [3-6]. The findings highlight the need for interface designs and decision-support mechanisms that promote critical evaluation and reduce the risk of overreliance on AI-generated health information.

Collectively, these contributions deepen our understanding of how GenAI shapes trust calibration in health information contexts, positioning learned dependency as a key mechanism through which habitual AI use influences users’ evaluative judgment.

We conducted 2 randomized controlled experiments to examine the impact of GenAI on human trust calibration. The experiments evaluated participants’ trust in AI-generated health information using a between-participants design with randomized assignment to conditions.

Our study was approved by the institutional review board (IRB) prior to data collection (IRB protocol: STUDY00002235). All participants provided informed consent before beginning the survey, covering the study’s purpose, estimated completion time, voluntary participation, and data storage procedures. No personally identifiable information was collected. Data were stored securely on the Qualtrics platform and were accessible only to the research team. The health information stimuli were obtained from the MIMIC-III (Medical Information Mart for Intensive Care) clinical database, a credentialed-access dataset, under a data use agreement held by the research team. The CHERRIES (Checklist for Reporting Results of Internet E-Surveys) checklist (Checklist 1) covers the online survey or experiment administered via Qualtrics for both experiments. Where responses differ between experiments, both are reported separately.

A total of 338 participants completed the first experimental study (Table 1 and Figure 3). The sample was nearly evenly split by sex (n=171, 50.59% male participants vs n=167, 49.41% female participants) and predominantly aged 18 to 30 (n=336, 99.41%). Most participants were identified as White (n=266, 78.70%), and 69 (20.41%) participants were identified as Hispanic or Latino. Educational attainment was relatively low, with 80.47% (n=272) having completed only high school. Reflecting the recruitment context, 97.93% (n=331) of participants were business or economics majors.

A total of 563 participants completed the second experimental study (Table 1 and Figure 4). The sample was predominantly male (n=439, 77.98%) and primarily aged 21 to 40 (n=530, 94.14%), with a mean age of 31.7 (SD 6.5) years. Most participants identified as White (n=538, 95.56%), and 23.62% (n=133) identified as Hispanic or Latino. Educational attainment was considerably higher than in experiment 1, with 72.82% (n=410) holding a bachelor’s degree and 22.20% (n=125) holding a master’s degree. Major fields of study were more varied, led by STEM (science, technology, engineering, and mathematics; n=180, 31.97%), health or medicine (n=110, 19.54%), and business or economics (n=109, 19.36%).

All statistical analyses were conducted using R version 4.2.1. Hierarchical multiple regression was the primary analytical approach, with variables entered in 3 blocks: block 1 included control variables (health care context, age, education, and sex), block 2 added the main effects (information accuracy, TVC, and learned dependency), and block 3 added the 2 interaction terms (information accuracy × learned dependency; TVC × learned dependency). Learned dependency was mean-centered prior to constructing interaction terms to reduce multicollinearity. Incremental model fit was assessed using ΔR² and the associated F-test at each block. Common method bias was assessed using the Harman single-factor test. To assess the psychometric properties of the multi-item constructs, a CFA was conducted for the trust scale and the learned dependency scale using the lavaan package in R. Model fit was evaluated using standard indices, including standardized root mean square residual, CFI, Tucker-Lewis Index, and RMSEA, and convergent and discriminant validity were assessed using standardized factor loadings, AVE, CR, and Cronbach α. Full CFA results, including factor loadings and fit indices, are reported in Multimedia Appendix 1.

To examine the effects of information accuracy, TVC, and learned dependency on trust in AI-generated health information, we conducted linear regression analyses. Table 2 presents the regression coefficients, and Table 3 summarizes hypothesis testing results. The results of experiment 1 (Table 2) were consistent with H1; information accuracy had a significant positive effect on trust (B=2.107, 95% CI 1.337-2.878; P<.001). The results show that participants reported substantially higher trust in accurate AI-generated information than in inaccurate AI-generated information, indicating that, on average, results are consistent with users being more sensitive to information accuracy when comparing correct and incorrect outputs. In contrast, TVC (highlighting) did not have a significant effect on trust (B=0.149, 95% CI −0.622 to 0.920; P=.70), providing no support for H2. This suggests that directing users’ attention through highlighting does not meaningfully influence their trust in AI-generated information. Supporting H3, learned dependency on GenAI was positively associated with trust (B=0.277, 95% CI 0.033-0.521; P=.03). Individuals with higher levels of dependency exhibited greater overall trust in AI-generated outputs. Among the control variables, most effects were not statistically significant. However, exposure to certain text conditions (health care context 4: B=−0.449; P=.01; health care context 5: B=−0.610; P<.001) was associated with lower trust, suggesting potential context-related effects.

To assess whether learned dependency influences trust calibration, interaction terms were included in the model (Table 2). Consistent with H4, the interaction between information accuracy and learned dependency was negative and significant (B=−0.399, 95% CI −0.695 to −0.104; P=.008). This finding indicates that higher dependency exhibits reduced sensitivity to information inaccuracy. In other words, highly dependent users are more likely to trust incorrect AI-generated information than less dependent users, reflecting a pattern consistent with diminished trust calibration, as described in the automation bias literature [20]. In contrast, the interaction between TVC, learned dependency, and accuracy was not significant (B=−0.023, 95% CI −0.618 to 0.573; P=.94), providing no support for H5 (Table 3). This suggests that attention-based interventions, such as highlighting critical information, do not mitigate the effect of dependency on trust. Overall, the results provide strong support for the role of learned dependency in shaping trust in AI-generated health information.

Experiment 2 (Table 2) replicated this pattern using the refined 4-item trust scale. Consistent with H1, information accuracy again had a significant positive effect on trust (B=0.203, 95% CI 0.115-0.290; P<.001). TVC again showed no significant effect on trust (B=0.013, 95% CI −0.074 to 0.101; P=.76), providing no support for H2. Supporting H3, learned dependency was positively and more strongly associated with trust than in experiment 1 (B=0.822, 95% CI 0.715-0.929; P<.001). The interaction between information accuracy and learned dependency was again negative and significant (B=−0.459, 95% CI −0.577 to −0.340; P<.001), providing further support for H4 and indicating that the trust-distorting effect of learned dependency replicated in a larger, more diverse sample with a strengthened accuracy manipulation. The interaction between TVC, learned dependency, and accuracy remained nonsignificant (B=0.142, 95% CI −0.095 to 0.379; P=.24), again providing no support for H5. Taken together, the cross-sample replication of H1, H3, and H4, alongside the consistent null findings for H2 and H5, strengthens confidence in the robustness of these effects (Table 3).

Although users generally trust accurate information more than inaccurate information, increased dependency weakens this distinction. Furthermore, text highlighting does not significantly influence trust or reduce overreliance among highly dependent users.

This study examined how characteristics of AI-generated health information and users’ behavioral reliance on GenAI influence trust in AI-generated health information. By integrating trust calibration theory and attention theory, our findings provide insights into how users evaluate AI-generated health information and how behavioral dependency may shape these evaluations.

First, the results suggest that information accuracy remains an important determinant of trust in AI-generated health information, which is consistent with trust calibration theory [27]. Across both studies, participants generally expressed greater trust in accurate AI-generated responses than in inaccurate ones. This finding indicates that users can calibrate their trust based on the quality of the information provided by AI systems.

However, the findings also reveal that learned dependency on GenAI plays a critical role in shaping users’ trust judgments. Individuals who reported higher levels of dependency on AI systems exhibited greater trust in AI-generated health information overall. More importantly, dependency weakened users’ ability to differentiate between accurate and inaccurate information. This pattern suggests that dependence on AI tools may reduce users’ ability to critically evaluate AI outputs, increasing the likelihood of overreliance on GenAI-generated information. From the perspective of human-automation interaction, this pattern is consistent with automation bias, in which users defer to automated systems even when the outputs may be incorrect, a pattern observed across both studies. These findings are consistent with theoretical mechanisms described in the automation bias literature [18,20,21] and the cognitive offloading perspective [22,31].

The results also provide insights into the effectiveness of the TVC in supporting users’ evaluation of AI-generated health information. Although attention theory suggests that directing users’ attention to critical information may improve information processing, the TVC intervention did not significantly affect trust calibration across both studies. This finding suggests that interface cues, such as text highlighting directed at specific portions of AI-generated information, may not be sufficient to change how users evaluate its credibility, including in experiment 2, where the TVC was substantially strengthened (doubled). Users who already depend heavily on AI systems may continue to rely on AI output regardless of such interface cues.

Taken together, these findings highlight an important behavioral tension in AI-assisted health information environments. Although users appear capable of recognizing differences in information accuracy, behavioral reliance on GenAI can weaken this evaluative process. As a result, dependency on AI tools may increase users’ vulnerability to trusting inaccurate AI-generated health information. These dynamics raise important considerations for the design and governance of GenAI systems in highly sensitive contexts such as health care, where incorrect information may have adverse consequences for health-related decision-making.

Several limitations should be considered when interpreting the findings of this study. First, while the inclusion of experiment 2 using a diverse MTurk population sample substantially improves upon the student-only design of experiment 1, MTurk samples themselves may differ from the broader general population in terms of digital literacy and online task experience. Future research should examine these relationships in clinical populations and among older adults with active health concerns. Second, the attention-based intervention across both experiments focused on text highlighting to direct users’ attention to key information. Importantly, the manipulation was strengthened in experiment 2 by highlighting 6 locations rather than 3, yet the null finding for H5 persisted, providing strong evidence that highlighting alone is insufficient to counteract dependency-driven trust distortion regardless of intervention intensity. Future studies should test more substantive cognitive interventions such as uncertainty indicators, confidence cues, or AI reasoning explanations. Third, the scenarios used in both experiments represent controlled and simplified representations of health information. Real-world health decision-making contexts often involve higher stakes, emotional factors such as fear, urgency, stress, or hope, and richer multimodal information formats including colors, images, and video, all of which may influence how individuals interpret and trust AI-generated information. Fourth, this study focuses on trust as the proximal outcome and does not measure downstream behavioral intentions or actual health decisions, which represent important next steps for this research. Fifth, no formal manipulation check was included for the accuracy manipulation, which was a deliberate design choice to avoid contaminating the trust measurement, but this limits independent verification of the manipulation’s effectiveness. Sixth, learned dependency was measured through self-report, which may be subject to social desirability bias. Future research should complement self-report measures with behavioral or longitudinal indicators. Seventh, both experiments used a controlled side-by-side presentation that differs from naturalistic health information seeking, limiting ecological validity. Eighth, the visual attention manipulation used bold text only; the null H5 finding should be interpreted as specific to this implementation rather than as evidence that all visual attention interventions would fail. Although bold-text highlighting is an established form of TVC in the information processing and multimedia learning literature [39-41], it represents a relatively subtle form compared to richer visual modalities such as images, color overlays, animations, or graphics. Accordingly, the conclusions of this study apply specifically to text-based health information delivery contexts and may not generalize to multimodal AI-generated health information that incorporates images, graphics, or other rich visual elements. Ninth, in experiment 1, the trust and learned dependency scales did not demonstrate discriminant validity (interconstruct correlation exceeded √AVE for both constructs), suggesting these constructs may not have been empirically distinguishable with the shorter scales used in that experiment. This concern was resolved in experiment 2 following CFA-based scale refinement, but it remains a limitation of the experiment 1 measurement model and should be considered when interpreting the H3 and H4 findings from that experiment.

Future research should pursue the following specific directions. First, future studies should test more substantive cognitive interventions beyond text highlighting, specifically uncertainty indicators that display the AI’s confidence or uncertainty alongside its output, to determine whether these engage users at a deeper evaluative level. Second, longitudinal designs tracking the same users over extended GenAI use would allow the examination of how dependency develops over time. Third, future studies should examine these relationships in clinical populations with active health concerns. Fourth, behavioral outcome measures such as verification-seeking should complement self-reported trust. Fifth, multimodal stimulus designs incorporating color, images, graphics, and variable typography should be tested to determine whether richer visual interventions overcome the null H5 finding. Sixth, future studies should examine how health literacy, medical knowledge, and digital literacy moderate the dependency-trust relationship and include more diverse and clinically relevant populations.

Across 2 controlled experiments with distinct populations, our findings provide robust empirical evidence that both characteristics of AI-generated information and users’ behavioral traits influence trust calibration in GenAI systems used for health information consumption. Information accuracy positively influences trust, and learned dependency consistently increases overall trust while reducing sensitivity to information inaccuracy. Critically, text highlighting failed to counteract dependency-driven trust distortion in either experiment, including experiment 2, where the visual intervention was substantially strengthened by highlighting incorrect information in 6 locations rather than 3. This pattern of null findings, replicated under conditions of stronger treatment, constitutes a substantive finding in its own right: the trust-distorting effect of learned dependency is robust and cannot be mitigated by surface-level attentional cueing alone. These findings highlight the importance of designing AI-enabled health information systems that promote appropriate trust calibration through more substantive cognitive interventions. As GenAI tools become increasingly integrated into digital health ecosystems, understanding how dependency shapes evaluative judgment will be critical for ensuring safe and responsible use of AI-generated health information [16,44].