When ignorance induces reliance: The role of epistemic knowledge about generative AI in changing health-related decision-making

Abstract

Today’s increasing and indiscriminate use of generative artificial intelligence to access science-related information is a serious cause for concern. Here, we investigated the role of knowledge about generative AI in individuals’ engagement with AI-generated science-related information. Specifically, we focused on the relationship between epistemic AI knowledge, regarding the features and construction processes that characterize AI output, and individuals’ likelihood of changing their dietary decisions following exposure to AI-generated content. Based on two consecutive online surveys with embedded performance tasks, Study I (n = 500) and Study II (n = 800) demonstrated that less epistemic AI knowledge and higher trust in generative AI significantly increase the likelihood of changing decisions and reasoning, often to align with the generative AI’s responses. People were also more likely to change their decision in the presence of information reliability cues such as two concordant generative AI responses, regardless of alignment with experts’ views or the presence of reliable sources.

Keywords

decision-making epistemic AI knowledge generative AI public engagement with science information

1. Introduction

Generative artificial intelligence (GenAI) increasingly serves as a source of scientific information, with reported usage doubling from 2023 to 2024 in multiple countries (Greussing et al., 2025). While excelling at providing personalized and easy-to-understand information, including science-related information (Biyela et al., 2024), GenAI also faces unique epistemic limitations. These embedded mechanisms and features, which can lessen and even impair the quality of the information generated, take different forms. For example, GenAI’s training on biased or unreliable data often results in biased or inaccurate outputs (Templin et al., 2024). GenAI is also prone to “AI hallucinations,” that is, linguistically viable and factually inaccurate information. As GenAI tools are designed to please their users, their misprompting—that is, incorrect, biased, or misleading prompt information—can easily lead the AI to generate misinformation (Klein-Avraham et al., 2026a; Koopman and Zuccon, 2023).

On the one hand, GenAI has the potential to augment users’ effective decision-making, speeding access to complex information and supporting information synthesis (Huynh, 2024; Klingbeil et al., 2024). In healthcare, for instance, AI can enhance accuracy in medical image analysis, ultimately supporting more accurate and efficient clinical decisions (e.g. Alruwaili et al., 2025). On the other hand, decision-supporting information generated by AI is exposed to the technology’s epistemic limitations, as mentioned. For instance, GenAI provides inaccurate diagnoses when peer-pressured (Shoval et al., 2025). Since decisions related to health, such as tumor identification or psychiatric assessment (Shoval et al., 2025), often involve high stakes and potential risks (Williams-Ceci et al., 2024), overreliance on AI recommendations in health-related contexts can be dangerous. By overreliance, we mean accepting unadvisable AI recommendations that misalign with experts’ views. Chiang and Yin (2022) showed that greater knowledge about the AI tool reduces overreliance among individuals capable of making decisions independently, particularly when the decision-makers’ expertise is superior to that of AI. This raises the question of whether an awareness of GenAI’s epistemic limitations can mitigate related risks of overreliance on AI advice. Put more generally, can awareness of GenAI’s epistemic limitations explain how individuals engage with science and health-related information generated by AI?

2. Theoretical background

Although several theoretical frameworks substantiate the expected association between knowledge about the technology and its use, the specific role of awareness of its epistemic limitations in shaping the use of the information it provides remains largely unaddressed. The reasoned action approach (RAA) (Fishbein and Ajzen, 2009) framework assumes that information or beliefs people possess direct their rational behavior, including decision-making. Although this framework considers people’s previous experiences, education, and demographic background as factors that guide their decision-making and behaviors, it focuses on individuals’ self-perceptions, that is, how these factors shape their beliefs about self-efficacy, social norms, and expected reactions from their close friends and family. The technology acceptance model (TAM) (Davis and Granić, 2024) builds on the RAA, focusing on individuals’ perceptions of the technology. According to TAM, perceived usefulness and perceived ease of use predict individuals’ attitudes and trust in new technology, ultimately influencing their intention to use it. The TAM, however, focuses its explanations on technology adoption and overlooks the potential effect of knowledge about the technology. The mental model matrix (MMM) framework explains that individuals’ “mental representation of how something works [guides] how they process information, anticipate future events, and interact with devices and tools in complex sociotechnical systems” (Borders et al., 2024: 75). Although this framework addresses the perceived capabilities and limitations of the technology, it aims to assess knowledge gaps between stakeholders, rather than explain individuals’ use of the technology’s output. Thus, although the literature supports the link between knowledge about GenAI’s mechanisms, capabilities, and limitations and the use of AI-generated information for decision-making, this issue is largely untheorized. Moreover, existing research lacks a structured approach to categorizing different dimensions of GenAI knowledge and understanding how they drive user behavior.

AI literacy, including knowledge of GenAI, is essential for effective use of the technology and for critically addressing AI-generated content (e.g. Long and Magerko, 2020). Although multiple AI literacy scales evaluate knowledge about AI, they are often based on self-reports and lack a systematic assessment of different types of AI knowledge (e.g. Carolus et al., 2023; Wang et al., 2022). Therefore, we drew inspiration for conceptualizing different types of AI knowledge from the OECD’s (2023: 10) PISA Science Framework, which differentiates between three types of scientific knowledge: (1) Content knowledge refers to the understanding of scientific facts, concepts, ideas, and theories about the natural world; (2) Procedural knowledge refers to the understanding of “the procedures that scientists use to establish scientific knowledge”; and (3) Epistemic knowledge refers to “the role of specific constructs and defining features essential to the process of knowledge building in science.” Therefore, we consider AI knowledge to encompass: content AI knowledge, that is, what AI and GenAI are and where they can be found in everyday life; procedural AI knowledge, which focuses on the workings and use of AI and GenAI technologies; and epistemic AI knowledge, which concerns the features and construction processes of AI outputs (Klein-Avraham et al., 2026b: 3–4). In other words, epistemic AI knowledge addresses awareness of AI’s and GenAI’s black-box nature, its susceptibility to social biases and stereotypes, and its potential to create inaccurate output. Measuring epistemic AI knowledge, therefore, captures individuals’ ability to critically evaluate AI-generated claims.

Besides knowledge about the technology, the literature identifies additional key factors in AI-supported decision-making and individuals’ adoption of AI technologies. It emphasizes trust in the technology, attitudes toward it, previous experience with it, and demographic background.

Trust in AI refers to one’s willingness to accept vulnerability based on positive expectations, encompassing perceived benevolence, integrity, and functionality. The more people trust a technology, the more inclined they are to adopt it (Choung et al., 2023) and accept its recommendations (Elder et al., 2024), even when the recommendations are incorrect (Buçinca et al., 2021; Klingbeil et al., 2024). Trust in AI and the acceptance of its recommendations are affected by its perceived reliability, which draws from the perceived accuracy and depth of the technology’s output (e.g. Elder et al., 2024).

Hence, individuals’ capability to evaluate the accuracy and trustworthiness of the information provided is crucial for informed decision-making (Forzani et al., 2022), especially in health-related dilemmas, given the high stakes involved (Williams-Ceci et al., 2024). The literature primarily differentiates between firsthand evaluation strategies, which assess the content itself, and secondhand evaluation strategies, which appraise the trustworthiness of the information source. Sourcing, that is, using available information about the origin of the claims, is a promoted practice for information evaluation (Barzilai et al., 2020). In the GenAI environment, this practice takes the form of meta-sourcing, that is, judging the appropriateness of the sources GenAI referenced in its response (Dabran-Zivan et al., 2026). A third strategy for evaluating information is corroboration, that is, comparing the provided information with that provided by other reliable sources (Barzilai et al., 2020). Triangulating information-evaluation strategies affords deeper and more accurate assessments of information reliability (Forzani et al., 2022). Nevertheless, the ability and willingness to engage in such evaluation strategies vary between people.

Attitudes toward AI encompass individuals’ emotional tendencies, positive and negative predispositions, beliefs, or opinions regarding AI (Bergdahl et al., 2023; Palm et al., 2025a). They are positively associated with trust in AI, and function as a key factor in the adoption of AI technologies (Bergdahl et al., 2023; Choung et al., 2023). Simultaneously, however, experience with the technology also positively affects attitudes toward it (Palm et al., 2025a) and trust in it (Schäfer et al., 2024). Gender, age, and education, too, explain trust and attitudes toward AI (e.g. Bergdahl et al., 2023; Palm et al., 2025b), which, as noted, explain AI adoption and the acceptance of its recommendations. Thus, younger and academically educated individuals tend to trust AI more and have more positive attitudes toward it (Gillespie et al., 2023). Males often report more positive attitudes toward AI (Bergdahl et al., 2023; Palm et al., 2025a), although not all studies corroborate this (e.g. Gillespie et al., 2023). Therefore, exploring the relationship between AI knowledge and the use of AI-generated information requires considering factors such as trust, attitudes, prior experience, and demographics.

The literature on the interplay between AI knowledge and trust, attitudes, and prior experience with GenAI is limited and inconsistent. While some studies link greater knowledge about AI with higher trust in the technology (e.g. Scantamburlo et al., 2024) and more positive attitudes toward it (Laupichler et al., 2024; Palm et al., 2025a), other studies depict negative interactions (e.g. Tully et al., 2025). Experience with AI, on the other hand, consistently demonstrates positive association with AI knowledge (Laupichler et al., 2024; Wang et al., 2022). It remains unclear, however, whether epistemic AI knowledge can explain how people use AI-generated information in decision-making, particularly regarding everyday science and health issues.

Against this backdrop, this article addresses two research questions:

RQ1: What are the relationships between individuals’ AI knowledge, especially epistemic AI knowledge, and their trust in GenAI, attitudes toward AI, and experience with GenAI tools?

RQ2: What factors, including epistemic AI knowledge, predict whether individuals will modify their decisions and reasoning when presented with AI-generated information?

This article describes two consecutive studies. Study I examined the relationships between general and epistemic AI knowledge, trust in GenAI, attitudes toward AI, and prior experience with GenAI tools. It also investigated whether individuals would revise their decisions about an everyday health-related issue following exposure to two authentic and agreeing GenAI responses. Study II addressed the limitations of Study I and tested its main findings. Hence, it examined the role of epistemic AI knowledge and trust in GenAI in the use of AI-generated information for health-related decision-making while allowing more nuanced insights into the role of information-evaluation strategies in such situations.

3. Study I: Does awareness of GenAI’s epistemic limitations explain how people use AI-generated information?

Method

An online survey was conducted in July–August 2023 among 500 participants who represent the adult internet user population in Israel according to age and education level (see Supplemental material: Supplementary 1). The sociotechnological context of Israel is characterized by a relatively high adult education level (51% tertiary education rate) and an 88% internet penetration rate (OECD, 2024; World Bank Open Data, 2024). At the time of data collection, GenAI tools underperformed in Hebrew compared to English, thereby guiding the technology selection for the study, as described below.

Measurements

The questionnaire comprised two parts. The first was co-designed for an international survey (see Greussing et al., 2025), including, inter alia, the following measurements (see Supplementary 2):

Attitudes toward AI were measured on a six-statement Likert-type scale (1 = “strongly disagree” to 5 = “strongly agree”, α = .769). An average index score was calculated for participants with three or more valid answers (n = 493, M = 3.235, SD = .752).

Trust in GenAI was measured on a 16-item Likert-type scale (α = .910). An average index score, ranging from 1 to 5, was calculated for participants with five or more valid answers (n = 483, M = 3.413, SD = .692).

Knowledge about GenAI was assessed through nine newly developed true/false statements inspired by Long and Magerko (2020), and after consulting with two AI experts (Greussing et al., 2025). The nine statements comprised five about AI in general and four about GenAI; six referred to content and procedural knowledge, and three measured epistemic knowledge. The data was transformed into two knowledge scores:

General AI knowledge (n = 500, M = 5.310, SD = 1.918), comprising the number of correct answers on all nine items, ranging from 0–9;

Epistemic AI knowledge (n = 500, M = 1.742, SD = 1.098), comprising the number of correct answers on the three epistemic items, ranging from 0 to 3.

Familiarity and experience with GenAI. Participants reported whether they had heard of or used five GenAI tools: ChatGPT, Bing Chat (currently Copilot), Bard (currently Gemini), Perplexity AI, and Robi Bot. These data were transformed into:

Accumulating experience with GenAI. For each technology tool, we recoded the reported use so regular use = 2 and using the technology once or twice = 1, summing the experience across the five GenAI tools. To accommodate the ordinal nature of the variable, the accumulating experience was then divided into three categories: no or almost no experience with GenAI tools (scores 0–1, n = 207, 41.4%), limited experience (scores 2–4, n = 187, 37.4%), and varied experience (scores 5–7, n = 38, 7.6%).

Demographic Variables included self-reported gender, age, and the highest level of education (Supplementary 1).

The second part of the survey included a performance task based on a vignette, introducing a friend who contemplated whether a detox juicing diet would be safe and beneficial for her health. Participants were asked two open-ended questions—what they would recommend their friend do and why—before and after being presented with AI-generated information. The AI-generated information comprised authentic, thus similar, responses from two GenAI tools, consequently affording both content evaluation and corroboration (Barzilai et al., 2020). ChatGPT’s and Bing Chat’s responses were used due to their performance in Hebrew at the time. Bing Chat also referenced three information sources in its response: healthline.com, mediacalnewstoday.com, and verywellfit.com. These sources, however, did not support meta-sourcing due to their unfamiliarity among locals. The two screenshots, both shortened versions of the original GenAI responses and designed to appear authentic, were similar in nature and based on the same prompt, derived from the vignette. Consistent with health experts’ views (e.g. NCCIH, 2019), the GenAI responses (see Supplementary 2) inter alia stated that there is no scientific evidence that supports the benefits of juice diets; these diets offer limited nutritional options, thus might be harmful and even dangerous; the human body carries out natural cleansing processes; it is better to maintain a healthy lifestyle and regular physical activity, and it is advised to consult an expert for personalized information.

Data analysis

We analyzed participants’ recommendations and justifications thematically, and whether their recommendations aligned with experts’ views. As participants repeatedly incorporated justifications in their recommendations, they were analyzed together. Two coders independently coded 10% of collected responses with an acceptable agreement rate for all variables (Krippendorff’s α over .7; Hayes and Krippendorff, 2007). The first author conducted all the content analysis. Based on the codebook (see Supplementary 3), we developed two variables:

Participants’ considerations. The 28 themes that emerged from participants’ answers were grouped into seven considerations (Figure 1).

Considerations aligning with themes in the GenAI responses. Three of the participants’ considerations aligned with themes mentioned in the GenAIs’ responses: (1) a recommendation to seek additional information from reliable sources (e.g. consult an expert), (2) a relevant physiological consideration (e.g. the human body does detoxification regularly and on its own), or (3) a relevant nutritional consideration that suggests either avoiding juicing diet, pointing to their unhealthy or inefficient nature, or generally suggesting that juicing diets are unadvisable.

Figure 1.

The changing frequency of participants’ considerations in Study I, and the alignment of recommendations with specific themes in the GenAI responses.

Statistical analysis

Spearman’s rank-order and Pearson’s correlations analyzed associations between variables. To determine the significance of post-exposure changes in recommendations and reasoning, we used McNemar’s test. Finally, testing whether demographic variables, accumulating experience with GenAI tools, attitudes toward AI, trust in GenAI, and general and epistemic AI knowledge explain how individuals used the AI-generated information (i.e. whether and how participants changed their recommendations and reasoning), we conducted binomial logistic regressions separately for each factor.

Results

Attitudes toward AI, trust in GenAI, and accumulating experience with GenAI tools positively correlated with general AI knowledge. However, epistemic AI knowledge was negatively associated with trust and attitudes toward AI, while positively associated with accumulating experience (Supplementary 4).

Prior to testing which factors best predict decision change following exposure to AI-generated information, we examined whether and how participants’ decisions had changed. Overall, recommendations and reasoning significantly changed following exposure to AI-generated information. This resulted from 119 (23.8%) recommendations shifting to align with experts’ views as well as the GenAI responses, and 27 (5.4%) shifting to misalign with them. As for changes in considerations (Figure 1), those aligned with the GenAI responses significantly increased by 27.4% (p < .001). Although participants were asked to consider the diet’s safety and efficiency, some addressed other considerations, for example, the diet’s costs or their own qualifications to advise on the matter. Post-exposure, such considerations decreased by 61.9%, alongside a 145.5% rise in relevant nutritional considerations. This suggests that when presented with AI-generated considerations that directly relate to the dilemma, individuals tend to forgo their previous unrelated considerations and embrace the AI-generated ones.

To observe which factors explain participants’ use of the GenAI recommendations, we divided the participants into four groups based on their pre- and post-exposure recommendations: (1) rejecting the GenAI advice: recommendation changed from aligning with experts’ views pre-exposure to misaligning post-exposure (n = 27, 5.4%); (2) unchanged aligning: recommendations aligned with experts’ views pre- and post-exposure (n = 240, 52.6%); (3) unchanged misaligning: recommendations misaligned with experts’ views pre- and post-exposure (n = 70, 15.4%); and (4) accepting the GenAI advice: recommendation changed from misaligning to aligning with experts’ views (n = 119, 26.1%). Overall, we included the responses of 456 (91.2%) participants, of whom we were able to identify aligning/misaligning recommendations pre- and post-exposure.

We assessed the independent effect of each measured factor on the likelihood of participants reaching their pre–post recommendation combination, using a binomial logistic regression (Table 1 and Supplementary 5). The tested factors included attitudes toward AI, trust in GenAI, epistemic and general knowledge of GenAI, gender, age, education, and accumulated experience with GenAI tools.

Table 1.

Factors predicting participants’ likelihood of changing their decision to one that aligns with experts’ views (Study I, n = 500).

	Aligned → Misaligned (n = 27)			Aligned → Aligned (n = 240)			Misaligned → Misaligned (n = 70)			Misaligned → Aligned (n = 119)
Variable	β(SE)	p	Odds ratio(95% CI lower-upper)	β(SE)	p	Odds ratio(95% CI lower-upper)	β(SE)	p	Odds ratio(95% CI lower-upper)	β(SE)	p	Odds ratio(95% CI lower-upper)
Attitudes	-.15 (.263)	.953	.985 (.588-1.684)	.086 (.125)	.494	1.089 (.852-1.393)	-.386 (.173)	.026	.680 (.484-.954)	.158 (.145)	.275	1.171 (.882-1.555)
Trust	-.014 (.291)	.963	.987 (.558-1.744)	-.079 (.138)	.566	.924 (.706-1.210)	-.364 (.185)	.049	.695 (.483-.999)	.375 (.163)	.021	1.455 (1.057-2.003)
General GenAI knowledge	-.103 (.100)	.304	.902 (.741-1.098)	.163 (.051)	.001	1.177 (1.066-1.301)	-.157 (.066)	.018	.855 (.751-.973)	-.066 (.055)	.237	.937 (.840-1.044)
Epistemic AI knowledge	-.053 (.178)	.766	.948 (.669-1.345)	.287 (.087)	< .001	1.332 (1.124-1.578)	-.177 (.116)	.129	.838 (.667-1.053)	-.227 (.096)	.018	.797 (.660-.962)

Aligned = recommendations aligned with experts’ views and the GenAI responses; Misaligned = recommendations misaligned with experts’ views; β = change in log odds. SE stands for “standard error,” and CI for “confidence interval.”

Participants who accepted the GenAI advice tended to have lower epistemic AI knowledge and greater trust in GenAI. The contributions of attitudes, general AI knowledge, demographics, and accumulating experience were nonsignificant. Participants who did not change their misaligning recommendation tended to have lower trust in GenAI, less positive attitudes toward AI, lower general AI knowledge, and less experience with GenAI tools. The contributions of epistemic AI knowledge and demographics were nonsignificant. These results mark trust as a positive predictor of accepting the GenAI recommendation. While general AI knowledge predicted maintaining the misaligning recommendation, epistemic AI knowledge predicted the recommendation’s change and acceptance of the GenAI advice. Participants who did not change their aligning recommendation, thus agreeing with the GenAI responses, tended to have higher education and higher general and epistemic AI knowledge. The contributions of trust, attitudes, experience with GenAI, gender, and age were nonsignificant. For the 27 participants who rejected the GenAI advice, the binomial logistic regressions yielded nonsignificant results for all the tested variables.

Interim discussion in Study I

The results from Study I correspond with and add to the literature: while the positive association between general AI knowledge and accumulating experience with GenAI tools echoes previous findings (Laupichler et al., 2024; Wang et al., 2022), the positive association between epistemic AI knowledge and accumulating experience provides additional insight. The findings also show that while general AI knowledge demonstrates positive associations with trust and attitudes toward GenAI, epistemic AI knowledge presents negative associations. This suggests that general AI knowledge fosters trust and positive attitudes, whereas a deeper understanding of its epistemic limitations undermines both. This explanation not only agrees with Long and Magerko (2020), who described how having more accurate knowledge about the abilities of AI can help people calibrate their trust in it, but might also mitigate the incongruent results reported in the literature (e.g. Palm et al., 2025a; Scantamburlo et al., 2024; Tully et al., 2025). Notwithstanding, the narrow scope of the AI knowledge scale (nine statements), especially for measuring epistemic AI knowledge (three statements), constituted a methodological limitation that we addressed in our second study.

The findings also suggest that individuals tend to accept GenAI recommendations and reasoning when making everyday health-related decisions, consistent with prior literature (Buçinca et al., 2021; Klingbeil et al., 2024). Among participants whose pre-exposure recommendations misaligned with those of GenAI, lower trust in the technology and less general knowledge about it predicted not accepting the GenAI recommendations. In contrast, higher trust in GenAI and less knowledge about its epistemic limitations predicted acceptance of the GenAI recommendations among this group. While a positive interaction between trust in AI and the adoption of its recommendations is expected (Elder et al., 2024; Klingbeil et al., 2024), the distinct role of epistemic AI knowledge reveals previously unexamined patterns. Specifically, the findings suggest that the likelihood of changing one’s decision following exposure to AI-generated information increases for individuals who are less aware of AI’s potential pitfalls as a source.

The meaningful and surprising insights from Study I necessitated a follow-up investigation to strengthen the findings’ robustness while addressing three methodological limitations: the knowledge scale had insufficient scope, as noted above, all the participants were exposed to the same combination of concordant GenAI responses, and the sources Bing Chat referenced did not support meta-sourcing by the local population. Study II, therefore, had two objectives. First, to verify whether epistemic AI knowledge and trust in GenAI predict how individuals utilize AI-generated information in everyday health-related decision-making. Second, to systematically remedy the identified methodological shortcomings. Accordingly, Study II employed a more comprehensive knowledge scale and incorporated strategic variation in GenAI responses, alternating between concordant and contradictory output, as well as between including reliable sources and sources of unknown reliability.

4. Study II: Do epistemic AI knowledge and trust in the technology predict changes in decision-making?

Study II focused solely on RQ2. It examined whether people change their decisions and reasoning following exposure to altering combinations of GenAI responses, and what role epistemic GenAI knowledge and trust in GenAI played in such situations.

Method

To accommodate the advancement in GenAI technologies and their adoption, Study II was conducted more than a year after Study I. The online survey was distributed in October–November 2024 among 800 participants who constituted a representative sample of the adult internet user population in Israel according to gender, age, and education level (see Supplementary 6). At that time, Israel was ranked ninth in The Global AI Index (n.d.), and the experience of internet-using adults with ChatGPT increased from 16% reporting having used it at least once in 2023 to 46% in 2024 (Greussing et al., 2025).

The research tool and descriptive statistics

Similar to Study I, the questionnaire comprised two parts. The first included the following measurements (Supplementary 7):

Attitudes toward AI were measured using the same scale as in Study I and similarly indexed (α = .706, n = 786, M = 3.757, SD = .678).

Trust in GenAI was measured using a shorter 11-item version of the Likert-type scale (Greussing et al., 2025) employed in Study I (α = .859, n = 781, M = 3.645, SD = .640).

Knowledge about AI was assessed using an augmented version of the previous scale. The new version was developed and validated in a joint effort, which yielded a two-axial, performance-based scale that measures different types and domains of AI knowledge. The development and validation process included four stages (MacKenzie et al., 2011): conceptualizing a theoretical framework, developing items and evaluating their relevancy through content validity, testing the models fit statistically using confirmatory factor analyses followed by scale refinement, and finally testing construct validity by confirming the new scale behaves as expected in relation to other theoretical constructs and theoretically expected group differences, using structural equation modeling, one-way ANOVAs, and independent-sample t-tests. The new scale includes 26 items, drawn from Greussing et al. (2025) or inspired by Touretzky et al. (2023; see also AI4K12, n.d.). To allow evaluation of different knowledge levels, the new items vary in difficulty. To facilitate true/false responses, some items were phrased in absolute terms, for example, by using the word “always.” The detailed account of the scales’ development and validation processes can be found in Klein-Avraham et al. (2026b); here, we focused on the relationship between AI knowledge and the use of AI-generated information. Based on the validated 26-item scale, the number of correct answers created four knowledge scores:

General AI knowledge included the number of correct answers across all 26 items, ranging from 0–26 (n = 800, M = 15.42, SD = 5.136).

Content AI knowledge included the number of correct answers to nine items that asked what AI and GenAI are and where we encounter them in everyday life, ranging from 0–9 (n = 800, M = 5.179, SD = 2.019).

Procedural AI knowledge included the number of correct answers to nine items that asked how AI and GenAI work and are used, ranging from 0–8 (n = 800, M = 5.333, SD = 1.9).

Epistemic AI knowledge included the number of correct answers to nine items that asked how the features and embedded mechanisms of AI and GenAI reflect on and manifest in their output, ranging from 0–9 (n = 800, M = 4.909, SD = 2.354).

Accumulating experience with GenAI was measured similarly to Study I, based on the familiarity and experience with five current GenAI tools: ChatGPT, Copilot, Gemini, Perplexity AI, and Claude (see Supplementary 7). Inspired by Greussing et al. (2025), the scale ranged between 1 (I am hearing about it here for the first time) and 6 (I use it daily), with 4 indicating “using the technology several times a month” and 5 “using the technology several times a week.” To support better comparison with Study I, we recoded the values, so reported regular use (i.e. values 4–6) = 2, and using GenAI once or twice (i.e. value 3) = 1. We summed the recoded values, then divided the scores into four categories, corresponding to those reported in Study I: no or almost no experience with GenAI tools (scores 0–1, n = 338, 42.3%), limited experience with GenAI tools (scores 2–4, n = 287, 35.9%), varied experience with GenAI tools (scores 5–7, n = 104, 13%), and extensive experience with GenAI tools (scores 8+, n = 71, 8.9%). The latter was added in response to the growing use of GenAI, in comparison to the data from Study I.

Demographic variables included self-reported gender, age, and the highest level of education (Supplementary 6).

The second part of the survey included a performance task that employed the same vignette as in Study I, with two modifications. First, we harnessed ChatGPT and Perplexity AI to provide two GenAI responses from each tool: one that aligns with experts’ views and one that does not. While ChatGPT was selected for its widespread use among the local population (Greussing et al., 2025), Perplexity AI was selected for its inherent list of information sources. Thus, we employed a 2 × 2 design, alternating between concordant and contradicting GenAI responses, and the inclusion of reliable information sources (e.g. two national public health organizations, the educational arm of a national and worldwide leading science institute, and the most visited online news site in Israel) versus sources of unknown reliability. It is important to note that Study II focuses on alternating signals of reliable information, thereby not including signals of unreliable information, such as adding a label stating that the content might be misleading (Wittenberg et al., 2025) or a label saying that the information presented contradicts the scientific consensus (Osborne and Pimentel, 2022).

Participants were randomly assigned to one of four groups, and each was presented with a unique combination of two authentic screenshots of the GenAI responses (Supplementary 7):

Group 1: Corroboration and meta-sourcing. Both GenAI responses aligned with experts’ views, and Perplexity’s included reliable sources, thus signaling information reliability through corroboration and meta-sourcing.

Group 2: Corroboration, no meta-sourcing. Both GenAI responses misaligned with experts’ views, and Perplexity’s included sources of unknown reliability. Because the sources’ level of reliability could not be determined, information reliability was signaled only through corroboration, although the information contradicted experts’ views on the matter.

Group 3: Meta-sourcing, no corroboration. Perplexity’s response aligned with experts’ views and included reliable sources, while ChatGPT’s response misaligned with experts’ views. This combination signaled information reliability through meta-sourcing.

Group 4: No corroboration nor meta-sourcing. ChatGPT’s response aligned with experts’ views, and Perplexity’s did not; the latter included sources of unknown reliability. This combination simulated missing signals for information reliability.

The second modification to the performance task involved the questions asking participants for their recommendations, pre- and post-exposure to the GenAI responses (Supplementary 7). In Study II, we transformed these into close-ended questions based on the thematic analysis in Study I. Participants’ recommendation options were: don’t do the diet; do or try the diet; seek additional information (i.e. from sources of unknown reliability); consult an expert such as a doctor or dietitian; consult other people such as previous users, family, friends; and other. The questions regarding the reasoning for the recommendation remained open-ended pre- and post-exposure.

Data analysis

We employed thematic content analysis of participants’ reasoning, identifying nine themes (Figure 3 and Supplementary 8), based on the coding scheme from Study I and corresponding to the GenAI responses in Study II. The nine themes formed the values for the variable “Participants’ considerations.” In line with the variable’s nominal nature, we used Cohen’s Kappa to assess agreement between two coders who independently coded 10% of the collected responses, with acceptable agreement rates across all variables (Cohen’s κ ≥ .689). The first author conducted all the content analysis.

Statistical analysis

McNemar’s tests assessed the significance of changes in participants’ recommendations and considerations, given the variables’ nominal nature. Binomial logistic regressions examined the role of AI knowledge scores, trust in GenAI, and other measured factors in participants’ change in recommendation post-exposure.

Results

Overall, the majority of participants’ recommendations changed significantly post-exposure (see Figure 2). Among Groups 1 and 2, recommendations that corresponded with the concordant GenAI responses significantly increased. Specifically, recommendations to avoid the diet increased by 100% (p < .001) when both GenAI responses aligned with experts’ views, and recommendations to do or try the diet increased by 100% (p < .001) when both GenAI responses misaligned with experts’ views.

Figure 2.

The changing frequency of participants’ recommendations in Study II.

Recommendations to consult sources of unknown reliability—that is, seek additional information and consult other people—decreased significantly, by 55.9% and 36.8%, respectively. This trend persisted among Groups 3 and 4, which were presented with contradicting GenAI responses. Among Group 3, “meta-sourcing no corroboration,” recommendations to seek additional information decreased by 72.2% (p = .004) and to consult other people decreased by 100% (p = .023). Among Group 4, “no corroboration nor meta-sourcing,” recommendations to seek additional information decreased by 63.6% (p = .007).

The findings highlight the added value of meta-sourcing, presenting additional significant changes for Groups 1 and 3. Specifically, among Group 3, “meta-sourcing no corroboration,” recommendations to avoid the diet increased by 104.6% (p < .001), suggesting that when corroboration is indecisive, the availability of reliable sources can support making a decision that aligns with experts’ views. Among Group 1, “corroboration and meta-sourcing,” recommendations to consult an expert decreased by 20.9% (p < .001). This suggests that when the information is signaled as reliable by both corroboration and meta-sourcing, the need for additional experts’ opinions decreases.

While most considerations did not significantly alter post-exposure (Figure 3), two did: reasoning that supports seeking additional information from sources of unknown reliability decreased by 46.2%, and considerations that do not address the dilemma directly (i.e. non-dietary, health, or information-related) decreased by 67.3%. These results reflect participants’ reduced need for additional information from sources of unknown reliability and suggest that when presented with considerations that directly relate to the dilemma, individuals tend to forgo their previous unrelated ones. Although the considerations “addressing AI or the GenAI responses” and “missing codable reasonings” (e.g. “as I mentioned before,” “because,” etc.) significantly increased as well, these trends are best explained by the research design, for example, the presentation of GenAI responses or the existence of previous responses. Accounting for the changes in considerations separately for each research group reveals no additional significant shifts, except for one: the appreciation of or preference for expertise or experts’ opinions significantly decreased among Groups 1 (by 23.2%, p = .008) and Group 2 (by 17%, p = .043).

Figure 3.

The changing frequency of participants’ considerations in Study II.

Finally, a binomial logistic regression examined which AI knowledge scores, if any, alongside trust, best predicted participants’ likelihood to change their recommendation. The compared models included alternating knowledge scores—that is, epistemic, procedural, content, and general—and trust, attitudes, gender, age, education level, and accumulating experience with GenAI tools. These covariates were included based on findings from Study I and existing literature linking demographic and attitudinal variables with technology adoption and use (see above).

As seen in Table 2, the analysis confirms that lower epistemic AI knowledge (β =-.105, p = .007) and higher trust in GenAI (β = .337, p = .031) significantly increase the likelihood of participants changing their recommendation following exposure to AI-generated information. The analysis also indicates that while accumulating experience with GenAI was not a significant predictor as a whole (Wald χ²₍₃₎ = 4.617, p = .202), a significant effect was found for individuals with limited experience, who are less likely to change their decision compared to individuals with no or almost no experience (β = -.581, p = .049). Other covariates, including the alternative AI knowledge scores, emerged as nonsignificant.

Table 2.

Epistemic AI knowledge, along with trust in GenAI, best explained the change in decisions following exposure to AI-generated information (Study II, N = 800).

Variable	Epistemic knowledge			Procedural knowledge			Content knowledge			General knowledge
Variable	β (SE)	p	Odds ratio(95% CI lower-upper)	β (SE)	p	Odds ratio(95% CI lower-upper)	β (SE)	p	Odds ratio(95% CI lower-upper)	β (SE)	p	Odds ratio(95% CI lower-upper)
AI knowledge score	-.09 (.040)	.024	.914 (.845-.988)	-.039 (.046)	.386	.961 (.879-1.051)	-.053 (.042)	.213	.949 (.873-1.031)	-.035 (.018)	.055	.966 (.932, 1.001)
Trust	.341 (.157)	.029	1.406 (1.035-1.911)	.440 (.150)	.003	1.553 (1.159-2.083)	.446 (.149)	.003	1.562 (1.167-2.093)	.397 (.151)	.009	1.487 (1.105-2.002)
Attitudes	-.089 (.138)	.517	.915 (.698-1.198)	-.086 (.138)	.534	.918 (.701-1.203)	-.099 (.137)	.47	.905 (.692-1.185)	-.087 (.138)	.528	.917 (.700-1.201)
Gender	.075 (.161)	.643	1.077 (.786-1.478)	.056 (.160)	.726	1.058 (.772-1.449)	.072 (.161)	.655	1.075 (.784-1.474)	.078 (.161)	.63	1.087 (.788-1.482)
Age	0 (.005)	.936	1 (.991-1.010)	.001 (.005)	.845	1.001 (.992-1.010)	.000 (.005)	.948	1.000 (.991-1.010)	.000 (.005)	.969	1.000 (.991-1.010)
Education level:
Up to high school with no diploma		.640			.471			.464			.609
High school with diploma	-.282 (.226)	.213	.754(.484-1.175)	-.298 (.226)	.188	.743(.477-1.156)	-.310 (.225)	.169	.733(.471-1.141)	-.287 (226)	.204	.750(.482-1.168)
Post-secondary (non-academic) diploma	-.183 (.264)	.488	.833(.497-1.397)	-.226 (.263)	.389	.797(.477-1.334)	-.225 (.262)	.390	.798(.477-1.334)	-.189 (.264)	.474	.828(.494-1.388)
Undergraduate academic degree	-.204 (.236)	.388	.816(.513-1.296)	-.266 (.236)	.260	.767(.483-1.217)	-.266 (.234)	.256	.767(.485-1.213)	-.207 (.238)	.384	.813(.511-1.295)
Postgraduate academic degree	-.404 (.287)	.159	.668(.381-1.172)	-.499 (.283)	.078	.607(.348-1.058)	-.492 (.283)	.082	.611(.351-1.064)	-.421 (.287)	.142	.656(.374-1.152)
Experience with GenAI tools:
No or almost no experience		.202			.114			.11			.196
Limited experience	-.121 (.185)	.515	.866(.617-1.274)	-.144 (.186)	.437	.866(.602-1.245)	-.146 (.184)	.428	.864(.602-1.240)	-.113 (.186)	.545	.893(.621-1.286)
Varied experience	-.614 (.291)	.035	.306(.306-.958)	-.701 (.289)	.015	.496(.282-.875)	-.699 (.287)	.015	.497(.283-.872)	-.6235 (.292)	.033	.536(.302-.950)
Extensive experience	-.038 (.296)	.898	.963(.539-1.720)	-.095 (.294)	.746	.909(.511-1.618)	-.085 (.295)	.772	.918(.515-1.636)	-.058 (.296)	.844	.943(.528-1.685)
χ²	33.617 (p < .001)			29.256 (p = .004)			30.059 (p = .003)			32.202 (p = .001)
R²_N	.059			.052			.053			.057

β = change in log odds. SE stands for “standard error,” and CI for “confidence interval.” R²_N = explained variance size (Nagelkerke R²).

5. Discussing both studies

Based on two consecutive online surveys integrating a performance task, Study I and Study II consistently demonstrated that individuals’ decisions regarding an everyday health-related issue significantly changed following exposure to AI-generated information, often aligning with the GenAI responses. Although neither study included an untreated control group as a comparative baseline for these results, this trend is consistent with, and is reinforced by, previous studies on AI-supported decision-making in general (e.g. Klingbeil et al., 2024) and health-related decision-making (e.g. Buçinca et al., 2021) in particular, which included an untreated control group in their design. Adding to the literature, the results showed that participants revised their reasoning following exposure to AI-generated information, and that the number of considerations that differed from those presented by the AI decreased post-exposure. A significant decrease also emerged in the need for additional information, especially from sources of unknown reliability, suggesting that GenAI responses were viewed as sufficiently informative and authoritative.

Across both studies, participants with lower epistemic AI knowledge and higher trust were significantly more likely to revise their decisions after exposure. Building on the well-established contribution of trust in technology (e.g. Klingbeil et al., 2024), these results highlight the critical role of epistemic AI knowledge when engaging with AI-generated information. These findings contribute to our understanding of AI literacy, underscoring the importance of systematically addressing epistemic AI knowledge. Although the literature considers AI literacy to encompass AI knowledge (Carolus et al., 2023; Wang et al., 2022) and frequently considers the epistemic limitations of GenAI (Templin et al., 2024), prior conceptualizations of AI knowledge do not account for the understanding of these epistemic limitations. Second, these findings suggest that greater epistemic AI knowledge may serve as a protective factor against overreliance on AI recommendations. Therefore, AI literacy interventions should explicitly cultivate epistemic AI knowledge, an aspect that receives little attention in AI literacy education programs (Atias and Mawasi, 2025; Lin et al., 2025). Cultivating epistemic AI knowledge is also expected to afford a more calibrated trust in GenAI and the information it provides (Klein-Avraham et al., 2026a), by supporting reflective trust, that is, reliance based on individuals’ informed belief about the technology’s trustworthiness (Ferrario et al., 2020).

Adding to the literature, Study II illustrated the role of information-evaluation strategies when engaging with AI-generated content. First, the results showed that when both GenAI responses were concordant, that is, corroboration signaled information reliability, participants’ decisions significantly shifted to align with the GenAI responses. Moreover, overreliance on GenAI responses occurred primarily when corroboration signaled information reliability. On the one hand, this suggests that overreliance can be minimized when the GenAI responses are contradicting, thus promoting a more systematic processing of the information (e.g. Buçinca et al., 2021). On the other hand, considering that individuals use similar prompts when comparing responses of different GenAI tools, and that the same prompts often elicit generally similar answers, individuals who similarly misprompt the GenAIs will probably receive generally similar—that is, corroborated—inaccurate responses.

Second, the results showed that when the GenAI response incorporated reliable sources, that is, meta-sourcing signaled at least one GenAI response as reliable, participants’ decisions significantly shifted to align with expert views. When both corroboration and meta-sourcing signaled information reliability, participants’ need to consult an expert significantly decreased. The availability of reliable sources seems to have encouraged decision-making that aligned with experts’ views, even when corroboration suggested the information is unreliable. These findings highlight the importance of both incorporating source attributions in GenAI responses and educating the public about meta-sourcing as an information-evaluation strategy.

The limitations of the research primarily stem from its specific cultural context and its design: the pre-experimental design in Study I and the lack of an untreated control group in Study II. These limitations restrict the generalizability and internal validity of the findings. Besides extending the cultural context, future studies can employ a true experimental design, including an untreated control group, thereby allowing causal inference. Another limitation concerns the timing of the trust and attitude measures, which were administered before the performance task, potentially affecting participants’ use of the AI-generated information. To mitigate this, the questionnaire and study procedure were uniform across all four randomized groups in Study II, thereby standardizing any potential priming effects across conditions. Future studies could also vary the timing of these measurements, administering them either before or after the performance task.

To conclude, this research delineates the critical role that epistemic AI knowledge plays in everyday health-related decision-making. In doing so, this research emphasizes the importance of recognizing GenAI’s epistemic limitations, especially for sound decision-making, as such technologies become increasingly embedded in everyday information-seeking behaviors.

Supplemental Material

sj-docx-1-pus-10.1177_09636625261442388 – Supplemental material for When ignorance induces reliance: The role of epistemic knowledge about generative AI in changing health-related decision-making

Supplemental material, sj-docx-1-pus-10.1177_09636625261442388 for When ignorance induces reliance: The role of epistemic knowledge about generative AI in changing health-related decision-making by Inbal Klein-Avraham and Ayelet Baram-Tsabari in Public Understanding of Science

Footnotes

Acknowledgements

We gratefully acknowledge the insightful input from Dr. Shakked Dabran-Zivan. Her thoughtful comments have significantly contributed to the development of this work.

ORCID iDs

Inbal Klein-Avraham

Ayelet Baram-Tsabari

Ethical considerations

The Behavioral Sciences Research Ethics Committee of the Technion—Israel Institute of Technology approved our survey (approval: 2023-031) on June 16, 2023 and (approval: 2024-084) on October 6, 2024.

Consent to participate

Respondents digitally approved their consent to participate following a written review and before starting the questionnaire.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was supported by Niedersächsisches Vorab, Research Cooperation Lower Saxony—Israel. Lower Saxony Ministry for Science and Culture (MWK), Germany [Grant No. 11- 76251-2345/2021 (ZN 3854)].

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Supplemental material

Supplemental material for this article is available online.

Author biographies

Inbal Klein-Avraham (PhD) is a Postdoctoral Fellow at the Faculty of Education in Science and Technology, Technion—Israel Institute of Technology. Her current research focuses on individuals’ knowledge about generative AI and their evaluation and use of the information AI generates.

Ayelet Baram-Tsabari (PhD) is a Full Professor of Science Education and Communication at the Technion—Israel Institute of Technology. Her research focuses on supporting effective science communication and the impact of science education on public engagement with science.

References

AI4K12 (n.d.) Available at: https://ai4k12.org/ (accessed 29 December 2024).

Alruwaili

Alshammari

Alhaiti

Elsharkawy

Ali

Elsayed Ramadan

(2025) Neonatal nurses’ experiences with generative AI in clinical decision-making: A qualitative exploration in high-risk nicus. BMC Nursing 24(1): 386.

Atias

Mawasi

(2025) Conceptualizing AI literacies for children and youth: A systematic review on the design of AI literacy educational programs. Computers and Education: Artificial Intelligence 9: 100491.

Barzilai

Thomm

Shlomi-Elooz

(2020) Dealing with disagreement: The roles of topic familiarity and disagreement explanation in evaluation of conflicting expert claims and sources. Learning and Instruction 69: 101367.

Bergdahl

Latikka

Celuch

Savolainen

Soares Mantere

Savela

, et al. (2023) Self-determination and attitudes toward artificial intelligence: Cross-national and longitudinal perspectives. Telematics and Informatics 82: 102013.

Biyela

Dihal

Gero

Ippolito

Menczer

Schäfer

, et al. (2024) Generative AI and science communication in the physical sciences. Nature Reviews Physics 6(3): 162–165.

Borders

Klein

Besuijen

(2024) Mental model matrix: Implications for system design and training. Journal of Cognitive Engineering and Decision Making 18(2): 75–98.

Buçinca

Malaya

Gajos

(2021) To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction 5(CSCW1): 1–21.

Carolus

Koch

Straka

Latoschik

Wienrich

(2023) MAILS—Meta AI literacy scale: Development and testing of an AI literacy questionnaire based on well-founded competency models and psychological change- and meta-competencies. Computers in Human Behavior: Artificial Humans 1(2): 100014.

10.

Chiang

C-W

Yin

(2022) Exploring the effects of machine learning literacy interventions on laypeople’s reliance on machine learning models. In: Proceedings of the 27th international conference on intelligent user interfaces, Helsinki, 22–25 March, pp. 148–161. New York, NY: Association for Computing Machinery.

11.

Choung

David

Ross

(2023) Trust in AI and its role in the acceptance of AI technologies. International Journal of Human–Computer Interaction 39(9): 1727–1739.

12.

Dabran-Zivan

Klein-Avraham

Baram-Tsabari

(2026) When the source is a bot: How people adapt their evaluation strategies to assess AI-generated content. PLoS One 21(3): e0345300.

13.

Davis

Granić

(2024) The Technology Acceptance Model: 30 Years of TAM. Cham: Springer.

14.

Elder

Canfield

Shank

Rieger

Hines

(2024) Knowing when to pass: The effect of AI reliability in risky decision contexts. Human Factors 66(2): 348–362.

15.

Ferrario

Loi

Viganò

(2020) In AI we trust incrementally: A multi-layer model of trust to analyze human-artificial intelligence interactions. Philosophy & Technology 33(3): 523–539.

16.

Fishbein

Ajzen

(2009) Predicting and Changing Behavior: The Reasoned Action Approach. Milton Park: Taylor and Francis.

17.

Forzani

Corrigan

Kiili

(2022) What does more and less effective internet evaluation entail? Investigating readers’ credibility judgments across content, source, and context. Computers in Human Behavior 135: 107359.

18.

Gillespie

Lockey

Curtis

Pool

Akbari

(2023) Trust in Artificial Intelligence: A Global Study. Brisbane, QLD, Australia: The University of Queensland; KPMG Australia.

19.

Greussing

Guenther

Baram-Tsabari

Dabran-Zivan

Jonas

Klein-Avraham

, et al. (2025) Exploring temporal and cross-national patterns: The use of generative AI in science-related information retrieval across seven countries. Journal of Science Communication 24(2): A05.

20.

Hayes

Krippendorff

(2007) Answering the call for a standard reliability measure for coding data. Communication Methods and Measures 1: 77–89.

21.

Huynh

M-T

(2024) Using generative AI as decision-support tools: Unraveling users’ trust and AI appreciation. Journal of Decision Systems. Epub ahead of 26 November. DOI: 10.1080/12460125.2024.2428166.

22.

Klein-Avraham

Jonas

Greussing

Taddicken

Baram-Tsabari

(2026a) Information foraging with generative AI: Usage patterns in Germany and Israel. Media and Communication 14: 11487.

23.

Klein-Avraham

Ston

Atias

Roll

Baram-Tsabari

(2026b) Measuring different types and domains of AI knowledge: Developing and validating a performance-based scale. Computers & Education 247: 105573.

24.

Klingbeil

Grützner

Schreck

(2024) Trust and reliance on AI — An experimental study on the extent and costs of overreliance on AI. Computers in Human Behavior 160: 108352.

25.

Koopman

Zuccon

(2023) Dr ChatGPT tell me what I want to hear: How different prompts impact health answer correctness. In: Proceedings of the 2023 conference on empirical methods in natural language processing (eds H

Bouamor

Pino

Bali

), Singapore, 6–10 December, pp. 15012–15022. Stroudsburg, PA: Association for Computational Linguistics.

26.

Laupichler

Aster

Meyerheim

Raupach

Mergen

(2024) Medical students’ AI literacy and attitudes towards AI: A cross-sectional two-center study using pre-validated assessment instruments. BMC Medical Education 24(1): 401.

27.

Lin

Dai

O-L

(2025) Constructionism in K-12 AI literacy education: A systematic review of pedagogical designs, student outcomes, and learning mechanisms. Journal of Educational Computing Research 63(7–8): 1748–1781.

28.

Long

Magerko

(2020) What is AI literacy? Competencies and design considerations. In: Proceedings of the 2020 CHI conference on human factors in computing systems, Honolulu, HI, 25–30 April, pp. 1–16. New York, NY: Association for Computing Machinery.

29.

MacKenzie

Podsakoff

(2011) Construct measurement and validation procedures in mis and behavioral research: Integrating new and existing techniques. MIS Quarterly 35(2): 293–334.

30.

NCCIH (2019) “Detoxes” and “Cleanses”: What you need to know. Available at: https://www.nccih.nih.gov/health/detoxes-and-cleanses-what-you-need-to-know

31.

OECD (2023) PISA 2025 science framework. Available at: https://pisa-framework.oecd.org/science-2025/ (accessed 8 November 2023).

32.

OECD (2024) Adult education level. Available at: www.oecd.org/en/data/indicators/adult-education-level.html (accessed 3 March 2026).

33.

Osborne

Pimentel

(2022) Science, misinformation, and the role of education. Science 378(6617): 246–248.

34.

Palm

Kingsland

Bolsen

(2025a) Trust, experience, and innovation: Key factors shaping American attitudes about AI. Public Understanding of Science 35(3): 295–312.

35.

Palm

Kingsland

Bolsen

(2025b) Framing affects support for the development of artificial intelligence in the United States. Science Communication 47(5): 730–752.

36.

Scantamburlo

Cortés

Foffano

Barrué

Distefano

Pham

, et al. (2024) Artificial intelligence across Europe: A study on awareness, attitude and trust. IEEE Transactions on Artificial Intelligence 6(2): 477–490.

37.

Schäfer

Kremer

Mede

Fischer

(2024) Trust in science, trust in ChatGPT? How Germans think about generative AI as a source in science communication. Journal of Science Communication 23(9): A04.

38.

Shoval

Gigi

Haber

Itzhaki

Asraf

Piterman

, et al. (2025) A controlled trial examining large Language model conformity in psychiatric assessment using the Asch paradigm. BMC Psychiatry 25(1): 478.

39.

Templin

Perez

Sylvia

Leek

Sinnott-Armstrong

(2024) Addressing 6 challenges in generative AI for digital health: A scoping review. PLoS Digital Health 3(5): e0000503.

40.

The Global AI Index (n.d.) Tortoise media. Available at: https://www.tortoisemedia.com/data/global-ai#rankings

41.

Touretzky

Gardner-McCune

Seehorn

(2023) Machine learning and the Five Big Ideas in AI. International Journal of Artificial Intelligence in Education 33(2): 233–266.

42.

Tully

Longoni

Appel

(2025) Lower artificial intelligence literacy predicts greater AI receptivity. Journal of Marketing 89(5): 00222429251314491.

43.

Wang

Rau

P-LP

Yuan

(2022) Measuring user competence in using artificial intelligence: Validity and reliability of artificial intelligence literacy scale. Behaviour and Information Technology 42(9): 1324–1337.

44.

Williams-Ceci

Macy

Naaman

(2024) Misinformation does not reduce trust in accurate search results, but warning banners may backfire. Scientific Reports 14: 10977.

45.

Wittenberg

Epstein

Peloquin-Skulski

Berinsky

Rand

(2025) Labeling AI-generated media online. PNAS Nexus 4: pgaf170.

46.

World Bank Open Data (2024) Available at: https://data.worldbank.org (accessed 3 March 2026).

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.07 MB

0.00 MB