We conducted this cross-sectional survey with an embedded randomized UI vignette experiment between May 22, 2025, and September 3, 2025. A 2-arm, between-subjects design compared a baseline generative AI chat interface with a safety UI bundle. Participants were randomized in a 1:1 ratio. Each participant viewed 2 scenarios in the assigned UI condition and completed outcome measures after each scenario. Analyses followed an intention-to-treat approach based on randomized assignment.

Participants were Chinese adults aged 60 years or older residing in mainland China. Recruitment was conducted through 3 channels, including community-based recruitment at senior activity centers or neighborhood community sites, outpatient clinic waiting area recruitment with on-site research staff, and online recruitment via WeChat groups that included older adults. To reduce exclusion of individuals with lower digital literacy, family-assisted or interviewer-assisted completion was permitted. Assistance mode was recorded for all participants and was included as a covariate and in sensitivity analyses. The survey was administered in simplified Chinese.

Inclusion criteria were age 60 years or older, residence in mainland China, ability to read Chinese or complete the survey with assistance, and provision of informed consent. Exclusion criteria were inability to provide informed consent, duplicate submissions identified using platform controls plus response-pattern review, and low-quality responses defined a priori as failure on an instructed-response attention check combined with an implausibly short completion time.

For online recruitment, the survey platform settings were used to reduce repeat submissions, and additional screening was performed using completion time, highly similar response patterns, and device or IP-based indicators available to the platform. For community and clinic recruitment, staff recorded whether a participant had already completed the survey on the same day, and responses were additionally screened post hoc for near-identical response patterns and implausible repetition across entries.

Randomization was implemented within the survey platform using built-in equal-probability assignment in a 1:1 ratio. Allocation was concealed until assignment. Condition assignment was maintained across both scenarios for each participant.

The intervention was defined as a practical safety UI bundle rather than separable component-level manipulations. The study was not designed to attribute effects to individual elements within the bundle, and inferences were limited to the bundled condition compared with baseline.

The bundle development process was pragmatic and reproducibility-oriented. The research team first identified interface elements that could be implemented without model retraining or retrieval-system changes, then selected low-salience source labels and a short uncertainty plus verification message that could fit within a mobile chat layout. Candidate wording was simplified to avoid technical terms, preserve readability for older adults, and avoid giving the impression that the generic source labels represented verified citations.

Two scenarios were developed to reflect common information needs among older adults while minimizing potential ethical and clinical risks. Scenario 1 addressed an everyday service or travel decision query. Scenario 2 addressed general well-being related to sleep quality and daytime fatigue. The generative AI answers were written conservatively and avoided diagnosis, medication dosing, or individualized clinical instruction.

Stimuli were presented as static mobile UI screenshots with large, high-contrast text. The interface layout, typography, spacing, icons, and answer length were held constant across conditions. Only the safety UI elements differed between arms. Participants could not interact with the chat interface itself, so the experiment captured responses to a controlled representation of an AI chat interface rather than the cognitive demands of real-time conversational use. Screenshots of the safety UI bundle for both scenarios are provided in Figures S1A-S2B and in Multimedia Appendix 1.

To reduce order effects, the presentation order of the 2 scenarios was randomized at the participant level by the survey platform. Participants remained in the same assigned UI condition regardless of scenario order.

After informed consent and eligibility screening, participants completed baseline measures including demographics, digital literacy, technology anxiety, and prior exposure to AI chat tools. Participants were then randomized to the baseline UI or safety UI bundle condition. Participants viewed the first scenario screenshot (assigned condition) and completed scenario-specific outcome measures, a brief manipulation check, and a comprehension quiz. Participants then viewed the second scenario screenshot (same assigned condition) and completed the same measures. The survey concluded with brief global attitude items and an optional open-ended question about desired interface features. Completion time and per-page dwell time were recorded automatically. Assistance mode (independent, family-assisted, and interviewer-assisted) was recorded for all participants.

Family-assisted or interviewer-assisted completion was allowed when participants had visual, motor, or digital-literacy barriers. During assisted completion, participants were shown the assigned screenshots whenever possible. If reading support was required, assistants were instructed to read the visible on-screen text and survey questions verbatim, including any source-label text or uncertainty/verification nudge displayed in the assigned condition, without explaining, emphasizing, or interpreting the interface cues or the intended purpose of the intervention. Assistance mode was recorded for all participants.

To strengthen outcome validity beyond self-report, a behavioral proxy for verification was included after each scenario. Participants were shown a small, nonemphasized expandable element labeled “来源信息 (可展开)” (source information [expandable]) placed below the outcome questions, not directly under the screenshot. Expanding it opened a brief neutral source information panel. The outcome was whether the participant expanded it at least once. This measure was interpreted as a low-friction proxy for curiosity about source information, not as evidence of substantive real-world verification such as cross-checking external websites, consulting clinicians, or comparing multiple information sources.

All Likert items were measured on a 7-point scale from “非常不同意” (strongly disagree) to “非常同意” (strongly agree), unless otherwise specified.

After each scenario, participants completed two brief manipulation check items: (1) a yes/no item assessing whether the interface displayed a source information prompt, and (2) a Likert item assessing whether they noticed “提示类信息” (tip-style information, including reminders to verify or consult professionals).

Verification intention was the primary outcome of the study. For each scenario, verification intention was measured using a 3-item study-created scale informed by prior work on information verification and AI reliance. The items assessed intention to cross-check information via other channels, preference for authoritative sources, and intention to consult professionals or experienced individuals before acting. Scores were averaged across scenarios and items; higher scores indicated stronger verification intention.

Reliance intention was a co-primary outcome. For each scenario, reliance intention was measured using a 3-item study-created scale informed by automation-trust and AI-adoption literature. The items assessed willingness to follow the advice, perceived sufficiency to adopt the advice directly, and likelihood of using the AI chat interface again for similar problems. A reliance intention score was averaged across scenarios and items; higher scores indicated stronger direct reliance intention.

The behavioral proxy expansion outcome was treated as a co-primary supportive measure. It was defined as expanding “来源信息 (可展开)” at least once, with scenario-level expansion reported as a secondary breakdown.

Digital literacy was measured using a 6-item study-created screener assessing self-rated ability to perform smartphone-based information search, app installation and updating, font-size adjustment, voice input use, privacy settings management, and handling common login or verification steps. Items were scored on a 5-point scale and averaged, with higher scores indicating greater digital literacy. Participants who completed the survey with family or interviewer assistance were not excluded from digital literacy analyses solely because assistance was provided. Their digital literacy data were retained as participant-reported baseline characteristics. Assistance mode was recorded separately for all participants and was included as a covariate in adjusted models to account for potential differences between independent, family-assisted, and interviewer-assisted completion.

Technology anxiety was measured using a 4-item study-created scale informed by prior technology-anxiety constructs. The items assessed nervousness with new technology, worry about making mistakes, avoidance of unfamiliar features, and perceived stress when learning technology. Items were averaged, with higher scores indicating greater technology anxiety.

Covariates included age, gender, education level, residence type, living arrangement, self-rated health, prior AI chat exposure, and assistance mode.

The survey included an instructed-response attention check. Duplicate prevention used survey platform settings plus post hoc response-pattern checks. Dwell time and completion time were recorded and used as quality indicators rather than stand-alone enforced thresholds to avoid disproportionate burden on older adults who may need more time or assistance. Low-quality responding was defined before analysis as failure of the instructed-response attention check plus either an implausibly short total completion time (<3 minutes) or a straight-line response pattern across all Likert outcome items.

The target sample size was set at 200 participants (approximately 100 per arm). This sample size was selected to support estimation of the between-group difference in the primary continuous outcome (verification intention) with adequate precision while remaining feasible across the recruitment channels. Using a 2-sample comparison framework with equal allocation, 2-sided α of .05, and a continuous outcome, a total sample of 200 provided approximately 80% power to detect a standardized mean difference of about 0.4 between arms. The sample size also supported adjusted regression models including prespecified covariates without unstable estimation. The study was powered for the primary main effect comparison; heterogeneity analyses were interpreted descriptively.

All analyses were conducted under intention-to-treat based on randomized assignment. Baseline balance between arms was described using standardized mean differences for key demographic and technology-related variables.

Primary analyses estimated the effect of condition (safety UI bundle vs baseline) on verification intention and reliance intention using linear regression with robust SEs, adjusting for prespecified covariates (age, gender, education, prior AI chat exposure, digital literacy, technology anxiety, and assistance mode). Results were reported as adjusted mean differences with 95% CIs. For verification intention, an unadjusted Cohen d was also reported descriptively.

The behavioral proxy expansion outcome was analyzed using logistic regression estimating the odds of expanding “来源信息 (可展开)” at least once, adjusting for the same covariates. Scenario-level expansion outcomes were summarized descriptively and analyzed in secondary models.

Secondary outcomes, including trust calibration index, perceived trustworthiness, usability/readability, and cognitive load, were analyzed using analogous regression models. For the comprehension score, distributional diagnostics were examined to determine whether a count model was necessary; because diagnostics did not indicate material model misspecification, the final analysis used linear regression with robust SEs, and effects were reported as adjusted mean differences on the 0-8 score scale.

Because assistance during completion could influence participants’ ability to complete the survey and could be associated with digital literacy, assistance mode was adjusted for in the primary models. In addition, a sensitivity analysis restricted to participants who completed the survey independently was conducted to assess whether the main findings were robust to exclusion of assisted completions.

A condition-by-digital literacy interaction term was examined to assess whether the intervention effect varied with digital literacy, and results were interpreted alongside uncertainty intervals.

Item-level missingness was below the predefined 5% threshold for all primary and secondary outcomes; therefore, complete-case analysis was used for the final analysis. Multiple imputation by chained equations was not applied.

Ethics approval was obtained from the Ethics Committee of Chengdu Gongyun Education and Management Research Institute (approval number GYLL25013). Although this study uses a randomized allocation design to present different user interface variants, it is classified as a cross-sectional vignette survey rather than a clinical randomized controlled trial involving a live intervention, and therefore clinical trial database registration was not required. Because the study involved survey responses and static vignettes rather than clinical intervention, it was classified as minimal risk. All participants provided informed consent before eligibility screening and randomization. Study data were deidentified before analysis, and no identifying clinical information was collected. No participant images or identifiable user screenshots were included in the manuscript or supplementary materials. Participants did not receive monetary compensation. The well-being scenario provided general nondiagnostic information only and advised professional consultation when appropriate. Family-assisted or interviewer-assisted completion was permitted for accessibility; research staff were not involved in outcome interpretation, but on-site staff were not formally blinded to randomized condition after assignment because screenshots were visible during assisted completion.

Item-level missingness was 0% for the study-created multi-item measures used in the primary and covariate-adjusted analyses. Internal consistency was good to excellent for the 2 primary self-report outcomes. Cronbach α was 0.890 for the 6-item verification intention score and 0.873 for the 6-item reliance intention score. Scenario-specific coefficients were also acceptable for verification intention and reliance intention, as shown in Table 1.

The study-created covariate scales also showed high internal consistency. Cronbach α was 0.959 for the 6-item digital literacy screener and 0.941 for the 4-item technology anxiety scale. For the secondary composite outcomes, α was 0.930 for perceived trustworthiness and 0.946 for usability/readability. The scoring rules, overall descriptive statistics, and reliability coefficients for the study-created multi-item measures are summarized in Table 1.

The descriptive distributions were consistent with the primary outcome pattern and did not indicate problematic floor or ceiling effects for the main self-report scales. Verification intention was higher in the Safety UI bundle arm, whereas reliance intention was similar between arms. Digital literacy and technology anxiety were also similar between randomized arms, supporting their role as covariates rather than major sources of baseline imbalance. Descriptive performance for the study-created scales and nonscale measures is reported in Table 2.

The comprehension outcome was scored as an 8-item objective quiz total and was not treated as a reflective psychometric scale because the items intentionally sampled distinct factual and caveat-related elements from the 2 vignettes. Cognitive load was assessed using one mental-effort rating per scenario, and source-information expansion was a binary behavioral proxy; therefore, Cronbach α was not interpreted for these measures.

Overall, the reliability and descriptive results support the internal consistency and scoring adequacy of the study-created verification intention, reliance intention, digital literacy, technology anxiety, perceived trustworthiness, and usability/readability scores. These findings provide measurement support for the adjusted intention-to-treat analyses of the primary and secondary outcomes.

A total of 236 individuals were approached across the 3 recruitment channels. Twenty-two individuals did not consent or did not start the survey. Of 214 who provided consent and started the survey, 14 were excluded before analysis because of ineligibility (n=6, <60 years), duplicate or near-duplicate entries (n=4), or low-quality responding defined as failing the attention check with an implausibly short completion time (n=4). The final analytic sample included 200 participants, with 100 randomized to the Baseline UI arm and 100 randomized to the Safety UI bundle arm (Figure 1).

Randomization produced broadly comparable arms (Table 3), although some baseline differences were observed. The mean age was 67.48 (SD 5.14) years in the Baseline UI arm and 67.65 (SD 5.18) years in the Safety UI arm. Female participants accounted for 54.0% (n=54) in the Baseline UI arm and 44.0% (n=44) in the Safety UI arm. College or higher education was reported by 14.0% (n=14) in the Baseline UI arm and 30.0% (n=30) in the Safety UI bundle arm. Prior use of AI chat tools was reported by 40.0% (n=40) in the Baseline UI arm and 41.0% (n=41) in the Safety UI arm. Standardized mean differences were largest for education and gender, so adjusted analyses retained prespecified covariates to reduce residual confounding.

The manipulation checks confirmed that participants perceived the intended UI differences. In the Safety UI bundle arm, 75.0% (n=75) reported noticing a source prompt, compared with 15.0% (n=15) in the Baseline UI arm. The “noticed tip” item also differed in the expected direction, with mean 5.20 (SD 1.13) in the Safety UI bundle arm and mean 3.26 (SD 1.15) in the Baseline UI arm (Table 4).

Verification intention was higher in the Safety UI bundle arm than in the Baseline UI arm. The mean verification intention score was 4.41 (SD 0.59) in the Baseline UI arm and 4.72 (SD 0.63) in the Safety UI bundle arm, corresponding to an unadjusted mean difference of 0.31 points on the 7-point scale (Cohen d=0.50). In covariate-adjusted regression, assignment to the Safety UI bundle was associated with a 0.293 point increase in verification intention (95% CI 0.128-0.457; P<.001), holding constant age, gender, education, prior AI chat use, digital literacy, technology anxiety, and assistance mode.

Reliance intention was similar between arms. The mean reliance intention score was 5.03 (SD 0.58) in the Baseline UI arm and 4.97 (SD 0.54) in the Safety UI bundle arm, with a small unadjusted difference of −0.06. In covariate-adjusted regression, the estimated association between the Safety UI bundle and reliance intention was −0.105 (95% CI −0.239 to 0.029; P=.13).

The behavioral proxy outcome showed a numerically higher proportion of participants expanding the optional “source information” element in the Safety UI bundle arm, but the adjusted estimate was not statistically definitive. Expansion at least once occurred in 27.0% (n=27) of Baseline UI participants and 42.0% (n=42) of Safety UI participants. In adjusted logistic regression, the odds ratio (OR) for expansion was 1.76 (95% CI 0.947-3.269; P=.07), suggesting a possible increase in source-information seeking, although the CI included the null.

Scenario-level expansion proportions were numerically higher in the Safety UI bundle arm, but these secondary binary outcomes were interpreted cautiously. In Scenario 1, expansion occurred in 21.0% (n=21) of participants in the Baseline UI arm and 34.0% (n=34) in the Safety UI bundle arm. In the adjusted logistic regression model controlling for age, gender, education, prior AI chat exposure, digital literacy, technology anxiety, and assistance mode, assignment to the Safety UI bundle was associated with higher odds of Scenario 1 expansion, although the CI included the null (adjusted OR=1.86, 95% CI 0.94-3.68; P=.08). In Scenario 2, expansion occurred in 18.0% (n=18) of participants in the Baseline UI arm and 31.0% (n=31) in the Safety UI bundle arm. The adjusted scenario-level model similarly favored the Safety UI bundle, again with uncertainty around the estimate (adjusted OR=1.92, 95% CI 0.95-3.88; P=.07). These scenario-level models were interpreted as secondary supportive analyses because the study was powered for the main continuous outcome rather than for scenario-specific binary expansion outcomes. Global attitude items were exploratory and showed no evidence that the safety UI bundle reduced willingness to use AI chat tools for low-risk everyday information seeking. Optional open-ended responses were sparse and were not treated as formal outcomes (Table 5).

Trust calibration, operationalized as the standardized difference between reliance and verification, favored the Safety UI bundle arm. The trust calibration index mean was 0.29 (SD 1.29) in the Baseline UI arm and −0.29 (SD 1.43) in the Safety UI bundle arm, indicating lower overreliance risk when the safety UI bundle was present. The adjusted association also favored the Safety UI bundle (β=−0.567, 95% CI −1.005 to −0.129; P=.01). Because the generative AI answers were intentionally conservative and did not include incorrect or hallucinated responses, the index should be interpreted as an exploratory balance between reliance and verification orientation rather than a definitive measure of correct trust calibration.

Comprehension scores were high in both arms and nearly identical. The mean comprehension total score was 6.32 (SD 1.08) in the Baseline UI arm and 6.33 (SD 1.14) in the Safety UI bundle arm. Distributional diagnostics did not indicate a need for a count model; therefore, the final adjusted analysis used linear regression with robust SEs. The adjusted association between arm and comprehension was small and not statistically distinguishable from zero (β=−0.132, 95% CI −0.428 to 0.163; P=.38).

Before outcome modeling, internal consistency was assessed for the 2 study-created multi-item secondary outcome composites. Cronbach α was 0.930 for perceived trustworthiness and 0.946 for usability/readability, supporting use of the averaged composite scores in the regression analyses.

Perceived trustworthiness showed a modest reduction in the Safety UI bundle arm. The mean perceived trustworthiness score was 5.39 (SD 0.70) in the Baseline UI arm and 5.20 (SD 0.61) in the Safety UI bundle arm. The adjusted β was −0.199 (95% CI −0.382 to −0.016; P=.03), indicating that adding safety cues improved verification orientation while slightly lowering perceived trustworthiness.

Usability/readability ratings were similar between arms, with mean 5.40 (SD 0.60) in the Baseline UI arm and 5.39 (SD 0.66) in the Safety UI bundle arm. The adjusted association was small and not statistically significant (β=−0.069, 95% CI −0.254 to 0.115; P=.46). Cognitive load tended to be lower in the Safety UI bundle arm, with mean 3.96 (SD 1.60) in the Baseline UI arm and 3.56 (SD 1.44) in the Safety UI bundle arm, although the adjusted estimate did not exclude zero (β=−0.346, 95% CI −0.762 to 0.071; P=.10; Figure 2 and Table 6).

The condition-by-digital literacy interaction for verification intention was near zero (β=−0.005, 95% CI –0.220 to 0.210; P=.96), suggesting that the Safety UI bundle effect on verification intention was broadly similar across the observed literacy range. To evaluate whether retaining participants who required family or interviewer assistance materially changed the observed intervention effect, we repeated the analysis among participants who completed the survey independently (n=134). The pattern remained consistent, with mean verification intention 4.50 (SD 0.58) in the Baseline UI arm and 4.70 (SD 0.66) in the Safety UI bundle arm. In the covariate-adjusted model restricted to independent completers, assignment to the Safety UI bundle was associated with higher verification intention (adjusted β=0.218, 95% CI −0.025 to 0.461; P=.08). The estimate was directionally consistent with the primary analysis, although the CI included zero.

Sensitivity analyses using complete-case data were identical to the main analytic sample because missingness did not exceed the prespecified threshold. Gender and education showed the most visible between-arm differences; because both were prespecified covariates in the primary adjusted models, these imbalances were already accounted for, and no separate post hoc adjustment was required.

The safety UI bundle was associated with higher verification intention on a 7-point scale, with an adjusted difference of about 0.29 and a moderate standardized effect. Reliance intention did not increase and showed a small, nonsignificant decrease, indicating that the bundle shifted users toward more verification without encouraging greater direct reliance.

The observed increase in verification intention with the safety UI bundle aligns with a growing body of human-centered AI and automation trust research demonstrating that interface cues, such as uncertainty indicators, source displays, or explicit verification prompts, can meaningfully shift user behavior toward more critical engagement and reduce blind reliance on AI outputs [3,4,21]. Prior studies have shown that while explanations and transparency features often increase users’ trust in AI systems, they do not always translate into greater verification or scrutiny of outputs [14-16]. In contrast, direct verification nudges, such as prompts to check sources or warnings about potential errors, have been found to be more effective at encouraging users to verify information rather than simply increasing their confidence in the system [3,17,18]. The effect size observed here is comparable to those reported for other UI interventions aimed at mitigating misinformation or promoting critical evaluation, such as warning labels and uncertainty communication strategies [17,22]. Notably, such interventions can shift intentions even when the underlying content quality remains unchanged, underscoring the value of interface-level design for safety [3,22,23].

This pattern supports trust calibration frameworks emphasizing that safe adoption of AI is not about simply increasing or decreasing trust but about aligning user reliance with system uncertainty and capability [4,24-26]. Overreliance, sometimes termed “automation bias,” is a well-documented risk when systems appear authoritative or present information without cues about limitations [24,27,28]. By explicitly framing uncertainty and encouraging cross-checking, the safety UI bundle likely counteracted authority cues that can lead to excessive deference to AI recommendations [17,19,29]. Similar effects have been observed in clinical decision support and navigation aids, where calibrated trust interventions reduced overreliance without undermining appropriate use [14,29,30]. The present findings extend this literature by demonstrating that such calibration can be achieved through simple UI modifications in a vignette setting with older adults, a group often considered at higher risk for overtrust due to lower digital literacy or unfamiliarity with automated systems [27,31].

The modest increase in behavioral verification proxies echoes prior research documenting gaps between stated intentions (self-report) and observable behaviors in digital environments [3,32,33]. Behavioral measures in survey-based experiments often yield smaller effects and greater variance than self-reported scales, particularly when the behavior is low-cost (eg, clicking to expand a source) and optional [23,32,33]. Stronger behavioral endpoints, such as actual web searching, time spent reviewing sources, or correctness-based incentives, tend to produce more robust effects but are less feasible in vignette studies [23,32]. Nonetheless, prior work suggests that even modest, noisy behavioral uplifts can support the directionality of intervention effects, especially when triangulated with self-report and manipulation checks [32,33]. The findings highlight both the promise and limitations of using lightweight behavioral proxies in survey experiments and underscore the need for richer behavioral tracking in future research [32,33].

This result contrasts with some studies on warning labels and uncertainty displays that report reduced comprehension or increased confusion, particularly among older adults or users with lower literacy [27,34,35]. In other cases, disclaimers or uncertainty cues have distracted users or reduced perceived clarity of information [34,35]. The present design may have avoided the pitfalls by employing large-font screenshots, a stable layout, and concise cues placed below the answer, features known to support comprehension and minimize cognitive load for older users [27,31]. While prior work sometimes finds that added UI elements increase cognitive load or reduce clarity, the nonworsening comprehension observed here suggests that conservative answer structure and scenario simplicity can help preserve understanding even when safety cues are introduced [27,31,34].

The observed reduction in perceived trustworthiness is consistent with research showing that explicit uncertainty communication can lower perceived competence or reliability of AI systems, even as it improves trust calibration and reduces overreliance [4,17,19]. Citation displays alone often increase perceived credibility and may inadvertently encourage overtrust; however, bundling citations with an uncertainty nudge appears to temper this effect by signaling to users that outputs should be verified rather than accepted at face value [16,17,19]. This aligns with literature distinguishing “trust” from “appropriate trust,” arguing that a small reduction in perceived trustworthiness may be acceptable, or even desirable, if it prevents overreliance on potentially fallible systems [4,19,29]. Similar trade-offs have been reported in studies of warning labels, misinformation interventions, and human-automation warning systems where increased vigilance comes at the cost of slightly diminished confidence in system outputs [17,22].

The findings are encouraging given gerontechnology and accessibility research emphasizing that additional UI elements can burden older adults, especially on small screens or in chat interfaces [27,31,34]. Prior studies have found that added explanations or complex provenance panels can reduce usability scores among older users or those with limited digital literacy [4,16]. The cues used here were intentionally short, visually distinct, and placed in predictable locations, a design approach similar to minimalistic “just-in-time” warnings shown to preserve usability while enhancing safety in HCI research [27,36]. If previous work has argued that safety cues harm user experience, the null usability differences suggest that careful cue design can maintain user satisfaction even as verification orientation is improved [27,36].

High rates of cue recognition bolster confidence in the internal validity of the experimental manipulation. Prior UI cue experiments have sometimes endured manipulation failure, where participants do not notice disclaimers or misunderstand uncertainty indicators, undermining inference about intervention effects [3,19]. Salience and comprehensibility of cues are especially critical for older populations who may miss subtle interface changes due to sensory or cognitive constraints [27,31]. The rates of cue recognition observed here appear comparable to those reported for effective labels and warnings in digital interfaces; design choices such as clear visual separation and concise language likely contributed to this salience without increasing burden [27,36].

While some studies suggest low-literacy users benefit more from explicit guidance or verification prompts [34,37], no meaningful moderation was observed here. Possible explanations include a restricted range of digital literacy within the sample, assistance reducing effective literacy gaps during completion, or the simplicity of the cue making it broadly effective across subgroups [34,37]. In contrast to literature showing strong moderation by education or eHealth literacy on intervention effects [37], the results suggest that well-designed safety UI bundles may be broadly applicable. Nevertheless, targeted tests in more diverse or lower-literacy samples remain warranted to confirm generalizability [37].

This study used randomized assignment to estimate the effect of a practical safety UI bundle on verification orientation in a generative AI chat interface, while holding prompt and answer content constant. The design targets older Chinese adults, a population with high potential benefit from generative AI assistance but potential vulnerability to automation bias, low digital literacy, and difficulty evaluating AI-generated information.

Several limitations should be considered. Outcomes were based primarily on self-reported intentions within a vignette screenshot setting rather than observed real-world behavior in deployed systems. Static screenshots may understate the cognitive demands of real-time generative AI interaction, where users formulate follow-up questions, interpret changing responses, and decide whether to verify under time pressure. The behavioral proxy was minimal and may not capture substantive verification actions such as checking external sources or consulting professionals. Because assisted completion was permitted for accessibility, some participants with visual barriers may have received the source labels and the uncertainty/verification nudge through verbatim oral reading by a family member or interviewer. For these participants, the intervention may have functioned partly as a verbally mediated warning rather than a purely visual UI nudge. Although assistance mode was recorded, adjusted for, and examined in sensitivity analyses, the study cannot fully isolate visual salience from orally mediated cue exposure among assisted participants.

The study tested a bundle rather than individual safety cues, so it cannot determine whether source labels, uncertainty messaging, or verification nudges drove the observed effects. The generic source chips were useful for avoiding false provenance but may have reduced ecological realism by removing the cognitive burden of evaluating real citations. Because the verification nudge explicitly recommended checking authoritative sources and consulting professionals, verification-intention items may have been susceptible to demand characteristics or social desirability bias. The 2 scenarios were intentionally low risk and conservatively written, which may have produced high comprehension scores and limited the ability to detect comprehension trade-offs during more complex, ambiguous, or error-containing AI responses. The answers did not contain false or hallucinated information, so reduced reliance may partly reflect undertrust of accurate content rather than trust calibration in the strict sense. The cross-sectional design cannot assess habituation, banner blindness, or long-term use. Assistance during completion may have altered subjective cognitive load or comprehension, and on-site staff were not formally blinded after assignment, creating a potential source of interviewer bias. Residual confounding from baseline imbalances, including education and gender, may remain despite covariate adjustment. Finally, the sample size was selected for the continuous primary outcome and was likely underpowered for the binary expansion outcome. Additionally, because several primary or supportive outcomes were tested and no formal alpha adjustment was applied, the possibility of inflated Type I error should be considered.

The findings suggest that simple, implementable safety cues may shift older users toward more verification without degrading comprehension or perceived usability, although perceived trustworthiness may decrease modestly. Design and development teams can treat verification nudges and transparent framing of uncertainty as promising directions for further testing, while recognizing that effects may depend on wording, timing, interface prominence, user characteristics, and whether real citations or interactive system behavior are present.

Future work should test interactive generative AI systems, real verification behavior outside the survey environment, repeated exposure over time, and factorial designs that disentangle source labels, uncertainty language, and verification prompts. Studies should also compare generic labels with real citations to determine whether added ecological realism changes cognitive load, comprehension, and verification behavior.

In this randomized static-vignette survey, a safety-oriented UI bundle in a generative AI chat interface was associated with higher verification intention among older Chinese adults, while reliance intention, comprehension, and usability/readability remained largely unchanged. The trust calibration index moved in a direction consistent with lower overreliance risk, but interpretation should remain cautious because the study did not include incorrect AI outputs and was not designed to isolate individual UI components.