The influence of exposure to AI-generated social media content on verification intention: the mediating effect of trust and the moderating effect of AI literacy

1 Introduction

Generative AI is a key component of modern digital communication. Today, with the advent of large language models, image generators, voice synthesis systems, video generators, and deepfake technologies, individuals, organizations, and automated systems can easily generate realistic content at an unprecedented scale and speed (Chesney and Citron, 2019; Dwivedi et al., 2023; Lim et al., 2023; Nightingale and Farid, 2022). This change is particularly significant on social media, where platforms are built to quickly share content, amplify it through algorithms, pass it along to others, and entertain with emotionally charged multimedia. Consequently, users are now faced with AI-generated social media content, such as captions, news summaries, comments, images, avatars, short videos, voiceovers, and synthetic people or events.

As the use of AI-generated content (AIGC) becomes more common, it offers opportunities for creativity, personalization, education, and effective communication. Meanwhile, it poses significant issues related to authenticity, source transparency, credibility, manipulation, and misinformation (Epstein et al., 2020; Lazer et al., 2018; Zhou and Zafarani, 2020). AI-generated content can be more than just false information, as it can also be designed to conceal the source, mimic human language, produce fabricated images, or give a realistic appearance. This is a feature that makes AIGC especially difficult for users to determine whether to believe, ignore, verify, or share information. The specific problem addressed here is whether routine encounters with AI-generated social media content are associated with lower content-specific trust and stronger intentions to verify that content, and whether AI literacy changes the extent to which users rely on trust when deciding whether to verify.

This is a decision-making process, and trust is at the heart of it. The original source, production process and evidential basis of information are not generally available to users in digital environments. Rather, they rely on source-, platform-, modality-, design-, popularity-, social endorsement-, and perceived credibility-related cues to make judgments (Flanagin and Metzger, 2000; Metzger et al., 2010; Sundar, 2008). These determinations become more difficult in the age of AI-generated social media content, as the material might include elements of professionalism and credibility despite being inaccurate, biased, or fabricated. Hence, the impact of AI-generated content on trust is significant to comprehend user behavior in the current information landscape.

A behavioral outcome of special interest is verification intention. It means that users are willing to fact check suspicious content, cross-check claims, examine whether content is AI-generated, and check sources before believing or sharing information (Pennycook and Rand, 2019; Pennycook et al., 2020). The importance of verification intention is because the sharing speed of social media is faster than the speed of careful consideration. Previous studies demonstrate that misinformation can be highly contagious when users are making intuitive judgments, instead of accuracy-based evaluation (Pennycook and Rand, 2021; Vosoughi et al., 2018). Verification becomes even more crucial in the realm of AI-generated content, where synthetic media can be extremely convincing and challenging to identify (Nightingale and Farid, 2022).

AI literacy may help users navigate this environment. AI literacy refers to the knowledge, skills, and critical competencies that enable individuals to understand, evaluate, use, and question AI systems and outputs (Long and Magerko, 2020; Ng et al., 2021). Users with higher AI literacy may be more aware of the limitations, biases, and risks of AI systems. They may also be better able to recognize that AI-generated content can appear credible while still requiring verification. Therefore, AI literacy may change the way trust influences verification intention: users with stronger AI literacy may be less likely to rely blindly on trust and more likely to verify content as a normal part of digital information behavior. Recent research in a non-Western educational setting also links digital literacy and trust with AI acceptance, but it does not examine verification of AI-generated social media content (Khoso et al., 2026a).

Existing evidence addresses three related but separate problems. Research on synthetic realism emphasizes detectability and perceived credibility (Nightingale and Farid, 2022), misinformation research examines accuracy attention and general discernment (Guess et al., 2020; Pennycook et al., 2020), and recent studies in non-Western educational settings examine AI-related trust, literacy, engagement, and acceptance (Khoso et al., 2025, 2026a; Zhang et al., 2025). These research streams do not jointly test whether exposure to AI-generated social media content is associated with verification intention through content-specific trust or whether AI literacy changes that trust-verification association among university-centered users in Jordan. The research gap therefore concerns a context-specific psychological pathway and its boundary condition, rather than the general existence of misinformation or the acceptance of AI tools.

The MAIN model serves as the main theoretical framework, as it helps to understand how users assess the trustworthiness of content that can be authored by humans, machines, or a combination of both (Sundar, 2008). The elaboration likelihood model offers another mechanism of processing: Trust can serve as a credibility shortcut, while AI literacy can be seen as a capacity to process content more systematically (Petty and Cacioppo, 1986). The focal credibility judgment is not a new model to be considered, but a theory of trust. In this integrated account, the study adds by separating AI-generated social media content from generic misinformation, testing the role of trust as an intervening statistical pathway between AI-generated social media content and the outcome, and offering evidence from a university-based Jordanian sample, as both an intervening variable and a boundary condition. The guiding research question is: What is the relationship between exposure to AI-generated social media content and intention to verify through trust, and what factors influence the relationship between AI literacy and the trust-verification relationship?

2 Literature review and hypotheses development

2.1 Exposure to AI-generated social media content

AI-generated social media content is digital content that is created, significantly altered, or improved using artificial intelligence systems and is published on social media platforms. This can feature text, images, audio, video, virtual influencers, synthetic faces, voice simulations, automated comments or AI-assisted posts (Floridi et al., 2018; Lim et al., 2023). Exposure to such content is the frequency and extent of user contact with AI-generated content during browsing, viewing, searching, communicating and interacting on social media.

There are a number of ways people can be exposed to AI-generated content. Firstly, algorithmic recommendation systems can sometimes focus on content that is visually appealing, emotionally impactful, or highly engaging, which might be AI-generated. Second, users might come across AIGC via peer sharing, influencer posts, ads, and features on the platform that promote remixing or automated editing. Thirdly, AI-generated content can be found outside of clearly AI-related contexts, which means that exposure is not intentional. For many social media users, the exposure to AIGC is common and hard to avoid through these pathways (Liu et al., 2025; Luo et al., 2019).

The MAIN model offers a helpful explanation for the effects of technologically mediated content on credibility judgements. Sundar (2008) states that when users assess digital information they use cues that involve modality, agency, interactivity, and navigability. The agency cue has particular significance in the context of AI-generated content, as users need to determine if the content is created by humans, AI, or a mix of the two.

When agency is unclear, users may experience uncertainty about the credibility, intention, and reliability of the content. Such uncertainty may shape both trust and verification intention.

Repeated exposure may coexist with either normalization or skepticism. Within the MAIN model, uncertain agency and limited provenance make source credibility harder to evaluate; awareness of synthetic manipulation may therefore be associated with weaker trust (Chesney and Citron, 2019; Guo and Wang, 2025; Hashmi et al., 2025). H1 accordingly examines exposure as a correlate of content-specific trust without assuming that every AI-generated post is misleading or that exposure has a uniform causal effect.

A direct association between exposure and verification intention is also theoretically plausible. In the MAIN model, ambiguous agency and synthetic modality can function as uncertainty cues, while the elaboration likelihood model suggests that perceived relevance and processing ability can support closer information evaluation (Petty and Cacioppo, 1986; Sundar, 2008). Empirical research shows that directing attention toward accuracy and strengthening media literacy can improve discernment (Guess et al., 2020; Pennycook et al., 2020), while work on psychological inoculation and image credibility links awareness of manipulation and source-related cues with more critical evaluation (Roozenbeek and Van Der Linden, 2019; Shen et al., 2019). Evidence from Indonesian undergraduates similarly situates reference validation within AI-related information evaluation (Taufikin, 2025). Although these studies do not establish that ordinary exposure itself causes verification, they justify testing a direct positive exposure-verification association alongside the trust-mediated pathway. Based on this reasoning, the study proposes the following hypotheses.

Theory-to-hypothesis link (H1): The MAIN model proposes that agency and modality cues guide digital-credibility judgments. Exposure to content with uncertain human, artificial, or mixed authorship can therefore be associated with content-specific trust. Trust theory defines that outcome as willingness to rely on information under conditions of uncertainty; it does not independently explain the exposure cue (Mayer et al., 1995; Sundar, 2008). Theory-to-hypothesis link (H3): The MAIN model identifies ambiguous agency and synthetic modality as potentially relevant information cues. The elaboration likelihood model explains how such cues may be associated with closer evaluation when users perceive the content as relevant and have sufficient processing ability. Together, these propositions support testing a direct exposure-verification association that is distinct from the statistical pathway through trust (Petty and Cacioppo, 1986; Sundar, 2008).

• H1: Exposure to AI-generated social media content significantly affects trust in AI-generated social media content.

• H3: Exposure to AI-generated social media content significantly affects verification intention.

2.2 Trust in AI-generated social media content

In general terms, the term “trust” refers to the expectation of being vulnerable to another party and that the other party will not take advantage of that vulnerability (Mayer et al., 1995). Users' trust in information environments is expressed in their trust that information is accurate, reliable, credible, and useful enough to make decisions (Flanagin and Metzger, 2000; McKnight et al., 2002). This is not just the trust from a platform or from a human source, but the trust of content generated by or influenced by computational processes that may not be known to a typical user of the platform.

There are a number of dimensions that impact the trust in AIGC. Users can assess the accuracy of the content, consistency with current knowledge, credibility of the platform or account, and transparency regarding AI usage. Explainability and perceived system reliability are also crucial as AI outputs can be hard for users to audit (Shin, 2021). Meanwhile, AI-generated content can mimic human writing and create compelling images and/or text that can cause users to perceive the content as credible (Nightingale and Farid, 2022).

Dual process theories and the elaboration likelihood model can be used to explain the relationship between trust and the intention to verify. If users trust the information they may use heuristic processing and feel less need to check for further information. Uncertainty might lead to more systematic processing, such as fact checking and cross-referencing, when trust is low (Kahneman, 2011; Petty and Cacioppo, 1986). Trust can thus hinder the intention to verify in the AIGC context when users believe that AI-generated content is credible, accurate, or professionally produced. In turn, decreased trust can lead to more verification, as users try to minimize uncertainty before accepting or disseminating information.

The impact of trust, however, can be not always but not necessarily bad. Trust can also motivate engagement with content, potentially prompting some to look at supporting evidence. However, it is likely that high trust will decrease the perceived need to verify in social media environment that is fast, attention limited, and heuristic. Accordingly, trust is expected to be a key predictor of verification intention.

Theory-to-hypothesis link (H2): Trust theory explains why confidence in an informational object can reduce the perceived need for additional assurance. The elaboration likelihood model specifies the corresponding processing proposition: trust may serve as a peripheral credibility shortcut, whereas lower trust is associated with more deliberate evaluation and verification intention (Mayer et al., 1995; Petty and Cacioppo, 1986).

• H2: Trust in AI-generated social media content significantly affects verification intention.

2.3 Verification intention

Verification intention is the tendency to perform actions to check the accuracy, authenticity, and source of information before accepting or disseminating. These behaviours include checking the source, checking multiple sources, using fact check tools, checking if content is AI-generated, and withholding sharing until they are confident enough (Pennycook and Rand, 2019; Pennycook et al., 2020). In the context of social media, verification intention is particularly relevant as users are likely to come across information in an emotionally charged way and at a fast pace.

Studies on misinformation show that people do not necessarily share misinformation with the intent of deceiving others, rather, their attention is drawn away from correctness towards social rewards, identity, novelty or emotional response (Pennycook and Rand, 2021). Providing accuracy prompts and media literacy interventions can make users more likely to think more carefully about information (Guess et al., 2020; Roozenbeek et al., 2020). The implications of these findings for AIGC are that AIGC can make fabricated information more realistic and scalable, thus increasing the need for verification.

The verification intention is not confined to the verification of the truth of a claim in the case of synthetic media. Users might also have to decide if an image was edited, a video was synthesized, a voice recording was simulated or a post was created by a boat or AI system. This wider type of verification needs motivation and ability. The proposed model emphasizes the role of intention because intention is an important precursor to behavior and is suitable for survey research on planned verification practices.

2.4 AI literacy

AI literacy is a collection of skills that allow people to grasp the concepts of AI, identify the use of AI, assess the outputs of AI, communicate with AI systems, and understand the ethical implications of using AI (Long and Magerko, 2020; Ng et al., 2021). It encompasses an understanding of the capabilities and limitations of AI, knowledge of how AI systems generate outputs, understanding of algorithmic limitations and bias, and the ability to critically evaluate information produced by AI.

AI literacy is not the same as digital literacy, as it is more specific to algorithmic processes, machine-generated outputs, automation, decision-making based on data, and interaction with AI. AI literacy enhances the ability of users to understand posts created by AI, assess if the content might be synthetic or manipulated, and understand that credible-sounding AI-generated content may not be accurate in social media settings. These skills are becoming crucial as AI tools become integrated into daily communication strategies. Research involving Chinese and other non-Western educational settings likewise connects generative-AI use with engagement, creativity, and digital-literacy development, although these studies do not directly measure verification of synthetic social media information (Khoso et al., 2025, 2026b; Zhang et al., 2025).

Elaboration likelihood model can be used to explain the moderating role of AI literacy. Users' processing of information systematically might be enhanced by a higher level of AI literacy, which could also contribute to their motivation. If a user is well-versed in AI, they might not take anything on faith and may check content even if they think it is credible. Low AI literacy users, on the other hand, might be more likely to trust or use surface cues due to their lack of knowledge or skills to assess AI-generated outputs. Hence, AI literacy is expected to influence the strength of the relationship between trust and intention to verify.

Previous research on media literacy and misinformation indicates that media-literacy-related skills can enhance the ability to discern and resist false or misleading information (Guess et al., 2020; Hobbs, 2010; Jeong et al., 2012). By analogy, AI literacy should enable users to go beyond passive acceptance and to more critically verify AI-generated content. Accordingly, the study proposes AI literacy as a moderator.

Theory-to-hypothesis link (H5): AI-literacy frameworks define the competence to understand AI systems, recognize their limitations, and evaluate generated outputs. The elaboration likelihood model identifies that competence as a processing-ability condition that can reduce reliance on trust as the sole credibility shortcut. AI literacy is therefore specified as a moderator of the trust-verification association, while its separate main effect captures general verification-related evaluative competence (Long and Magerko, 2020; Ng et al., 2021; Petty and Cacioppo, 1986).

• H5: AI literacy moderates the relationship between trust in AI-generated social media content and verification intention.

2.5 Trust as a mediating mechanism

Mediation is the process of explaining how and why an independent variable affects a dependent variable by an intervening variable (Preacher and Hayes, 2008). Trust is viewed as a mediator in this study because exposure to AI-generated social media content is expected to impact users' trust judgments, which in turn will impact verification intention. Such a sequential logic is similar to that of the integrative model of trust, which proposes that trust is built from the assessment of cues, and then influences behavioral intentions (Mayer et al., 1995).

The mediating role of trust is also in line with the MAIN model and information-processing theories. AIGC gives users agency, modality, and source cues that impact credibility and trust. After trust judgments are made, they can either decrease or increase the need for verification. Therefore, exposure could have an impact on the intention to verify not only by increasing vigilance, but also by the psychological route of trust.

Trust offers a concise explanation for the indirect association between exposure and verification intention: if greater exposure is associated with lower trust, and lower trust is associated with stronger verification intention, the product of these two negative relationships is a positive indirect association. A remaining direct association may reflect vigilance not captured by trust. These are theoretically specified statistical pathways rather than evidence of temporal ordering or causation.

Theory-to-hypothesis link (H4): The MAIN model accounts for the exposure-to-trust component through agency and modality cues, while trust theory and the elaboration likelihood model account for the trust-to-verification component through reliance and credibility-related processing. Combining these propositions supports an indirect statistical exposure-trust-verification association; it does not establish temporal sequence or causation (Mayer et al., 1995; Petty and Cacioppo, 1986; Sundar, 2008).

• H4: Trust mediates the relationship between exposure to AI-generated social media content and verification intention.

3 Conceptual framework

The conceptual framework involves exposure to AI-generated social media content, trust in such content, intention to verify, and AI literacy. Exposure is the focus predictor, verification intention is the outcome, and trust is an intervening statistical pathway. AI literacy is defined as both a direct predictor of the intention to verify and a moderator of the relationship between trust and intention to verify. The inclusion of the literacy main effect allows the interpretation of the interaction beyond the general verification tendency of the respondents.

The estimated framework consists of four direct structural paths: exposure to trust, exposure to verification intention, trust to verification intention, and AI literacy to verification intention. It also contains the indirect exposure-trust-verification association and the trust × AI literacy interaction predicting the intention to verify. This specification aligns with the paths described in Tables 1, 2 and shifts the focus of AI literacy from exposure-verification to trust-verification relationship.

Table 1. Bootstrapped structural path coefficients (all p < .001).

Table 2. Bootstrapped mediation and moderation results (all p < .001).

AI literacy is viewed as a relatively existing evaluative competence that can influence the strength of the trust–verification relationship and can also directly influence the trust–verification intention. The cross-sectional data would suggest that the sequence of events is exposure or trust changes literacy prior to the development of intention to verify literacy, which would be a logical error to specify literacy as a mediator.

There was no redundancy in the general media awareness, information overload, AI anxiety, platform trust, and prior misinformation experience. Each of these could reasonably be considered an omitted antecedent, competing explanation, or confounder. They were left outside the specified model to maintain a focused test of exposure, trust, and AI-specific evaluative competence; their omission consequently limits the interpretation of the observed associations.

Figure 1 presents the conceptual model, showing the direct paths from exposure to trust and verification intention, the direct association between AI literacy and verification intention, the indirect statistical pathway through trust, and the moderating role of AI literacy on the trust-verification association. The added AI-literacy main-effect pathway is highlighted in yellow.

4 Methodology

4.1 Research design

This study adopts a quantitative, cross-sectional survey design. A quantitative approach is appropriate because the study tests theoretically derive relationships among latent constructs. A cross-sectional design is suitable for collecting perceptions, attitudes, and intentions at a single point in time from social media users. Although cross-sectional data cannot prove causality, they are widely used in communication, information systems, and digital media research when the aim is to test structural relationships among constructs (Hair et al., 2022).

Both partial least squares SEM (PLS-SEM) and covariance-based SEM (CB-SEM) are defensible for the present reflective measurement model. CB-SEM would be particularly suitable for strict confirmatory theory testing and global covariance-model fit. PLS-SEM was selected because the study prioritizes explaining variance in verification intention, simultaneously estimates direct, indirect, and latent interaction terms, and applies adapted indicators to an emerging AI-generated-content context; bootstrapping also permits inference without requiring multivariate normality (Hair et al., 2019; Hair et al., 2022). The large sample and the existence of established scales do not, by themselves, make CB-SEM inappropriate; the choice reflects the study's prediction-oriented objective and interaction specification rather than claim that PLS-SEM is the only valid approach.

4.2 Population, sampling, and data collection

The accessible study population comprised digitally active, university-affiliated adults in Jordan who reported encountering AI-generated content on platforms such as TikTok, Instagram, Facebook, YouTube, X/Twitter, LinkedIn, or comparable digital environments. Recruitment was university-centered and nonprobabilistic, with respondents reached across northern, central, and southern Jordan. Accordingly, the sample should be described as a convenience sample drawn largely from Jordan's higher-education environment rather than as a representative sample of all Jordanian social media users.

The questionnaire was circulated through university-associated channels and social media networks connected with the northern cluster centered on Yarmouk University, the central cluster centered on the University of Jordan, and additional university-associated and social media networks in southern Jordan. Respondents participated through convenience-based recruitment rather than individual selection from a defined sampling frame. The available study records do not document the number of people approached, the number who viewed or declined the questionnaire, precise fieldwork dates, recruitment scripts, or any incentive arrangements. Consequently, neither a conventional response rate nor incentive status can be verified or reported as an established procedural fact.

After screening incomplete responses, identifiable duplicate responses, unusually rapid responses, and straight-lined responses, the final dataset contained 1,150 valid questionnaires. A transparent sample-size benchmark for the four predictors in the verification-intention equation (exposure, trust, AI literacy, and trust × AI literacy) was calculated using an omnibus multiple-regression F-test and the noncentral F distribution. With α = .05, target statistical power of .80, and a small effect size of f² = .02, the required sample is approximately 602 participants. At N = 1,150, the corresponding estimated power is .982, and the minimum detectable omnibus effect at .80 power is approximately f² = .0104. This benchmark supports sample-size adequacy for the specified equation but does not establish representativeness or guarantee equivalent power for every indirect effect.

The retained records document only the final analytical sample (N = 1,150); they do not preserve the initial number of submitted questionnaires, the number excluded for incomplete answers, duplicate records, very fast completion, or straight-line responding, or the sequence and overlap of those exclusions. Here, incomplete responses refer to missing questionnaire information, duplicates to apparently repeated submissions, very fast responses to unusually short completion times, and straight-line responses to invariant or near-invariant answer patterns. However, the original duplicate-identification rule, time threshold, straight-line cutoff, and stage-specific counts were not retained and therefore cannot be verified or reconstructed. Only the final analytic N can be reported as an established screening outcome.

Geographical coverage improves contextual breadth within the accessible university-centered sample, but it does not convert convenience recruitment into probability sampling. The findings are therefore most applicable to digitally active respondents connected to Jordanian higher-education settings and should not be generalized without qualification to non-students, older adults, or the national social media population.

4.2.1 Sample coverage and demographic profile

Table 3 summarizes the demographic and contextual profile of the 1,150 respondents. Although northern, central, and southern regions were represented, the sample was concentrated in higher-education settings: 74.8% reported a bachelor's-level academic category, and the remaining respondents were classified within diploma or postgraduate categories. The table therefore describes a geographically diverse but university-centered convenience sample rather than a nationally representative cross-section of social media users.

Table 3. Demographic characteristics of the sample (N = 1,150).

4.3 Measurement instrument

All constructs were assessed with multi-item measures using five-point response scales. Exposure items used frequency anchors ranging from 1 = Never to 5 = Very frequently, while trust, AI literacy, and verification-intention items used agreement anchors ranging from 1 = Strongly disagree to 5 = Strongly agree. Existing measures were adapted by mapping each source item to the relevant construct definition, replacing generic references to information or technology with AI-generated social media content, and incorporating platform-relevant examples such as posts, images, videos, captions, and voiceovers. Conceptual correspondence, wording clarity, and contextual relevance were checked against the construct definitions, and Table 4 reports the retained item wording and source studies. The available records do not document an independent expert panel, translation/back-translation, cognitive interviews, or a formal pilot test; accordingly, none of those procedures is claimed. The reported loadings, reliability coefficients, and AVE provide post-collection psychometric evidence but do not substitute for a documented pretest.

Table 4. Measurement scale, item sources, outer loadings, and indicator VIF.

The table combines the measurement items, item-level literature sources, outer loadings, and indicator VIF values for the retained reflective indicators. AIGSMC was measured using a frequency scale (1 = Never, 5 = Very frequently), whereas TRUST, AIL, and VERIFY were measured using agreement scales (1 = Strongly disagree, 5 = Strongly agree).

Exposure to AI-generated social media content was assessed using items that measured how often the respondents came across posts, videos, visuals, captions, voiceovers, and other AI-generated content on social media. This operationalization is suitable since the construct is about repeated contact with AI-generated content and not about an attitudinal statement.

The five AIGSMC indicators were study-specific, literature-informed adaptations rather than verbatim adoption of an existing validated AI-generated-content exposure scale. Kim and Chen (2015) measured social-media information use through frequency items ranging from Never to Very frequently; their seven-point frequency logic was adapted here to a five-point format and a different content domain. Dwivedi et al. (2023) informed the generative-AI context, while Epstein et al. (2020) informed the conceptual distinction between AI-produced artifacts and human-generated content. Neither of the latter studies supplied a validated social-media exposure scale or the exact wording used in Table 4.

At the item level, AIGSMC1 adapts the general frequency-of-contact logic to AI-generated social media content, and AIGSMC3 applies that logic to repeated appearance in a social-media feed. AIGSMC2 was newly worded to capture encounters with specific formats (videos, images, posts, captions, and voiceovers); AIGSMC4 was newly worded to capture exposure across text, image, video, and audio formats; and AIGSMC5 adapts the frequency approach to named social-media platforms. Thus, the Kim and Chen (2015) study supports the frequency-response structure, whereas Dwivedi et al. (2023) and Epstein et al. (2020) support the generative-content domain rather than ownership of the individual item wording.

Content coverage was justified by mapping the indicators to four facets of the stated construct: overall encounter frequency, recurrent feed exposure, diversity of AI-generated formats, and platform-specific encounters. This conceptual mapping supports content relevance but is not equivalent to documented expert review, cognitive interviewing, or pilot validation, none of which can be verified from the available records. The observed AIGSMC loadings (.812–.849), composite reliability (.917), and AVE (.690) provide post-collection psychometric evidence rather than independent proof of prior content validity.

Trust in AI-generated social media content was assessed using items related to credibility, reliability, confidence, and usefulness. These indicators reflect the willingness of users to trust AI-generated social media content as a credible information source.

AI literacy was operationalized using items that assessed the respondent's understanding of AI systems, how to evaluate AI-generated output, awareness of bias and limitations, and ethical understanding of the use of AI. This multi-dimensional approach aligns with AI literacy scholarship, which focuses on not just technical knowledge, but also critical and ethical assessment of AI-generated content.

Verification intention was assessed using items that tapped into respondents' intention to check sources, fact-check posts or videos that seem suspicious, compare content with other sources, and verify before believing or sharing content created by AI. These items directly relate to the dependent variable and are related to verification behavior as opposed to past behavior.

The verification-intention items describe concrete actions, such as checking sources, comparing information, and fact-checking suspicious content, rather than asking respondents to endorse a general claim that verification is morally desirable. Standardized response anchors and documented written informed consent provide procedural transparency, but the available records do not establish anonymity assurances, a social-desirability scale, a marker-variable test, or an observational verification measure. Social desirability and common-method bias therefore remain possible and are acknowledged as limitations.

The final PLS-SEM assessment was based on outer loadings and indicator-level collinearity, in addition to the measurement scales and item-level sources. As displayed in Table 4, the retained indicators were above the recommended loading threshold of .70, and all VIF values were lower than the conservative threshold of 3.3, which means that the indicators were reliable and there was no problematic multicollinearity. Literature sources from which each item was adapted are also reported in the table.

Note. Loading = standardized outer loading; VIF = outer model variance inflation factor. All loadings exceed .70 and all VIF values remain below 3.3, supporting indicator reliability and the absence of critical collinearity.

4.4 Data analysis strategy

Data analysis was performed using SmartPLS and following the standard two-stage PLS-SEM procedure: measurement model assessment and structural model assessment. Data was screened for missing values, outliers, response quality and distributional characteristic prior to model estimation. While multivariate normality was not a requirement for PLS-SEM, it was important to screen for good data quality and transparent reporting.

Indicator loadings, Cronbach's alpha, composite reliability, rho_A, average variance extracted (AVE), and discriminant validity were used to assess the measurement model. Indicator loadings were interpreted in relation to the recommended loading of .70, and Cronbach's alpha, rho_A, and composite reliability were expected to be >.70. AVE values greater than .50 were considered as convergent validity. The heterotrait-monotrait ratio (HTMT) was applied to evaluate discrimination. If the HTMT is below .85 or .90, depending on the conceptual similarity, the constructs are empirically discriminated (Hair et al., 2022; Henseler et al., 2015).

Collinearity diagnostics, path coefficients, coefficients of determination, effect sizes, model fit indicators, and bootstrapping were used to evaluate the structural model. Variance inflation factor values were checked to make sure that the paths were not distorted by collinearity. Bootstrapping with 5,000 resamples was used to calculate t-values, p-values and confidence intervals for testing hypotheses. Reported R² values were for trust and intention to verify and f² values for the contribution of each predictor.

Mediation was assessed by estimating the indirect association from exposure to verification intention through trust. The indirect estimate was considered statistically significant when its 95% bootstrapped confidence interval excluded zero (Preacher and Hayes, 2008). The direct exposure-verification path was retained to distinguish partial from complete statistical mediation. Moderation was examined using a trust × AI literacy interaction term predicting verification intention while also including the direct AI-literacy main effect. Simple-slope estimates at low, average, and high AI-literacy levels were used to interpret the interaction (Henseler, 2017).

Demographic and contextual data such as region, gender, age group, educational level, university affiliation, use of social media daily and main social media platform were collected to describe the sample and to contextualize the empirical findings. The primary structural model was based on the constructs as theoretically specified, in order to maintain parsimony and avoid model inflation.

5 PLS-SEM results

This section reports the SmartPLS algorithm and bootstrapping output for the final sample of 1,150 valid responses collected from Jordanian social media users. The model included exposure to AI-generated social media content (AIGSMC), trust in AI-generated social media content (TRUST), AI literacy (AIL), verification intention (VERIFY), and the interaction term AIL × TRUST. The results are presented in accordance with the two-stage PLS-SEM reporting procedure, beginning with the measurement model and followed by the structural model and hypotheses testing.

The results are reported in a sequence that mirrors PLS-SEM best practice. First, the adequacy of the measurement model is established to show that the constructs were measured reliably and validly. Second, the structural model is evaluated to determine the explanatory power of the model and the strength of the hypothesized relationships. Finally, the bootstrapping results are used to make hypothesis-testing decisions regarding the direct, mediating, and moderating effects.

5.1 Measurement model

The reflective measurement model showed acceptable convergent validity and internal consistency. Each indicator loading was greater than the recommended .70: AIGSMC ranged from .812 to .849, AIL ranged from .802 to .853, TRUST ranged from .836 to .878, and VERIFY ranged from .855 to .887. All Cronbach's alpha, rho_A and AVE values were ≥.70 and ≥.50, respectively, which is considered to be acceptable for reliability and convergent validity (Hair et al., 2022). The detailed construct reliability and convergent validity statistics are shown in Table 5.

Table 5. Construct reliability and convergent validity.

The heterotrait-monotrait ratio (HTMT) was used to examine the discriminant validity. All HTMT values were below the conservative .85 threshold, which is indicative of discriminant validity between the latent constructs (Henseler et al., 2015) as shown in Table 6.

Table 6. Discriminant validity using HTMT.

The HTMT value between verification intention and trust was .474. This moderate degree of association is theoretically expected because trust is specified as a predictor of verification intention, yet .474 remains well below the conservative .85 criterion for discriminant validity (Henseler et al., 2015). The value therefore indicates related but empirically distinguishable constructs rather than problematic overlap. Moreover, HTMT reflects the magnitude of cross-construct indicator relationships and should not be interpreted as the direction or size of the structural path coefficient.

Figure 2 shows the graphical representation of the measurement model, including the reflective indicators and their loadings on the latent constructs that were retained.

5.2 Structural model

Path coefficients, explained variance, effect sizes and model fit indicators were used to assess the structural model. The estimated model accounted for 12.1% of the variance in TRUST and 34.8% of the variance in VERIFY. The coefficient of determination results are given in Table 7. The overall fit of the estimated model was also acceptable (SRMR = .058 and NFI = .955) as indicated in Table 8.

Table 7. Coefficient of determination.

Table 8. Model fit summary.

Figure 3 displays the structural model with the estimated relationships between the latent constructs, the direct paths and the interaction term that was used to test for moderation.

5.3 Hypotheses testing using bootstrapping

The bootstrapping results confirmed that the path coefficients were statistically significant and in the theoretically expected directions for the reduced research model. Exposure to AIGC was negatively associated with TRUST and positively associated with VERIFY. TRUST was negatively associated with VERIFY, indicating that higher trust in AI-generated social media content corresponded to lower verification intention. AIL also positively predicted VERIFY, and the positive interaction coefficient showed that AI literacy significantly weakened the negative TRUST → VERIFY relationship. The full bootstrapped structural path coefficients are presented in Table 1.

As reported in Table 2, the bootstrapped mediation output showed a positive and statistically significant indirect association from AIGSMC to VERIFY through TRUST [β = .094, t = 8.460, p < .001, 95% CI (.074, .117)]. This estimate is consistent with the specified model because exposure was negatively associated with trust (β = −.348), and trust was negatively associated with verification intention (β = −.270). Conditional estimates further indicated that the negative TRUST-VERIFY association was strongest at low AI literacy (β = −.393) and weakest at high AI literacy (β = −.147). These are model-based associations and do not establish temporal or causal direction.

To facilitate interpretation of the moderation effect, Figure 4 displays the simple slope analysis for the relationship between trust in AI-generated social media content and verification intention at low, medium, and high levels of AI literacy.

The final hypothesis-testing decisions for H1 through H5 are summarized in Table 9.

Table 9. Bootstrapped hypotheses testing summary (all p < .001).

Overall, the bootstrapping results indicate support for the proposed structural relationships in the Jordanian sample.

6 Discussion of the findings

The discussion follows the five hypotheses and interprets the findings as an integrated pattern of statistical associations. Particular attention is given to the distinction between the direct exposure-verification association, the indirect association through trust, and the conditional role of AI literacy; none of these cross-sectional estimates is interpreted as proof of causation.

6.1 Summary of the empirical pattern

The bootstrapped PLS-SEM estimates form a coherent pattern that is consistent with the proposed conceptual framework. Exposure to AI-generated social media content was negatively associated with trust in such content (β = −.348, t = 13.121, p < .001), trust was negatively associated with verification intention (β = −.270, t = 10.824, p < .001), and exposure had a positive direct statistical association with verification intention (β = .293, t = 11.943, p < .001). The estimated indirect association through trust was also significant (β = .094, t = 8.460, p < .001), and the trust × AI literacy interaction predicted verification intention (β = .123, t = 5.668, p < .001). Collectively, these findings are consistent with both a trust-related pathway and a separate verification-oriented association, but they do not demonstrate causal effects.

6.2 Exposure to AI-generated social media content and trust

The negative correlation between exposure to AI-generated social media content and trust aligns with worries regarding dubious authorship, potential manipulation, and unclear sources. This interpretation is consistent with the “liar's dividend” account, which highlights the difficulty of trusting digital evidence when it is possible to be aware of synthetic media (Chesney and Citron, 2019), and with research on deepfakes and media credibility (Guo and Wang, 2025; Hashmi et al., 2025). The observed association should be understood as a relationship, as the cross-sectional design does not allow to conclude that the exposure was the cause of the decrease in trust, whether the skeptical respondents noticed the AI-generated content more often, or whether another factor could have been responsible for both.

The finding also supports the relevance of source and agency cues in AI-mediated communication. The MAIN model proposes that people use technological and source-related cues when evaluating digital information (Sundar, 2008; Sundar et al., 2007). When authorship may be human, artificial, or hybrid, agency cues can be difficult to interpret. The observed exposure-trust association is therefore consistent with uncertainty about authorship and authenticity, particularly because AI-produced media can appear highly convincing (Nightingale and Farid, 2022). This explanation remains theoretical because the proposed uncertainty process was not measured directly.

6.3 Trust and verification intention

The negative relationship between trust and verification intention was supported, suggesting that as trust in AI-generated social media content increases, users are less inclined to verify the content. This finding is theory-important as it places trust as a cognitive short-cut in the processing of information online. If users find content credible, reliable, or valuable, there might be a diminished need to verify the source, compare it to other information, or determine if it's AI-generated. This is in line with social and heuristic models of online credibility assessment that indicate that users often use cues and shortcuts when processing online information, rather than systematically verifying it (Flanagin and Metzger, 2000; Metzger et al., 2010).

The observed pattern is compatible with the elaboration likelihood model and dual-process theory: greater trust may coincide with relatively less effortful credibility judgments, whereas verification intention reflects a willingness to evaluate information more systematically (Kahneman, 2011; Petty and Cacioppo, 1986). Research on misinformation also indicates that accuracy attention and cognitive reflection are relevant to the evaluation of questionable claims (Pennycook and Rand, 2019, 2021; Pennycook et al., 2020). Accordingly, the negative trust-verification association suggests a potential practical concern, but it does not establish that trust caused respondents to verify less.

6.4 Direct effect of exposure on verification intention

The relationship between exposure to AI-generated social media content and intention to verify is positive and direct, and statistically different from the trust pathway, consistent with a vigilance explanation. Those who saw more AI-generated content also expressed greater plans to verify the source or authenticity. This is in line with the research about synthetic realities, reference validation, and awareness of manipulative information (Das et al., 2025; Roozenbeek and Van Der Linden, 2019; Taufikin, 2025). But the increase in vigilance was not directly measured and the cross-sectional data does not reveal a precautionary attitude as a result of exposure.

The direct and indirect routes are not mutually exclusive. One association is between exposure and increased intention to verify, which may be due to uncertainty about authorship or synthetic cues. A second association is between increased exposure and decreased trust, and decreased trust and increased verification intention. The product of the two component paths in the latter sequence is a positive indirect estimate since both are negative. These interpretations are in line with separate vigilance-related and trust-related correlates of verification, but do not imply causation.

6.5 Mediating role of trust

The indirect effect was significant suggesting that the relationship between exposure to AI-generated social media content and intention to verify was statistically mediated by trust in the model specified. There was a positive indirect association between exposure and verification intention, through the mediation of trust. This pattern is consistent with the trust theory, which relates credibility judgments to the willingness to trust information when it is uncertain (Mayer et al., 1995; McKnight et al., 2002). This is a statistical mediation, however, as the constructs were measured at the same time, it does not prove a temporal or causal mechanism. The indirect estimate brings the discussion of verification to the fore, as it identifies trust as one theoretically plausible correlate between exposure and reported intention to verify.

This interpretation is in line with credibility research on the reliability of information and the trustworthiness of information sources (Appelman and Sundar, 2016; Lucassen and Schraagen, 2011). It is about self-reported intention and not about observed verification behavior and should therefore be considered as an explanatory statistical pattern and not as proof that exposure caused verification behavior. The association between direct exposure and verification was not eliminated by the inclusion of trust in the model, and thus the mediation was partial. The direct coefficient (β = .293) and the indirect coefficient through trust (β = .094) are thus complementary parts of the positive total coefficient (β = .388). The direct component is congruent with a vigilance-related interpretation, while the indirect component is congruent with lower content-specific trust. These mechanisms are conceptually separable, but the data cannot measure vigilance separately, nor determine the causal order of these mechanisms.

6.6 Moderating role of AI literacy

The significant interaction suggests that the relationship between trust and verification intention was moderated by AI literacy. The association between trust and verification was negative for all respondents, with the strongest association for respondents with low AI literacy (β = −.393) and the weakest for respondents with high AI literacy (β = −.147). This trend is like the one that emerged when respondents were asked about their intentions to verify; the more AI-literate were less reliant on trust alone. The finding is a conditional association, not a causal one that improving AI literacy led to a change in verification.

This interpretation aligns with the frameworks of AI-literacy which focus on understanding, assessing and critically engaging with AI systems and their products (Long and Magerko, 2020; Ng et al., 2021). It also fits in well with the research on media-literacy in terms of discernment and misinformation (Guess et al., 2020; Jeong et al., 2012). In the current model, AI literacy serves as a measured cognitive resource that is related to increased verification intention and decreased negative relationship between trust and verification intention.

The moderation pattern is also potentially relevant to education. The lower AI literacy respondents seemed more likely to report intentions to verify that were related to their trust, while the higher AI literacy respondents were less likely to show a difference between their intentions to verify based on their trust. It aligns with the work of Khoso et al. (2025), Liu et al. (2025), Obadă et al. (2026), and Zhang et al. (2025) on AI literacy, critical engagement, and AI use in non-Western and communication contexts, respectively. The effect of education on verification behavior is an issue for intervention research.

6.7 Overall contribution to literature

Overall, the study contributes to research on AI-generated content, misinformation, and verification by showing that, within this university-centered Jordanian convenience sample, greater reported exposure was associated with lower trust and stronger verification intention. The results extend discussion of deepfakes and digital credibility to routine AI-generated social media content (Chesney and Citron, 2019; Nightingale and Farid, 2022; Tandoc et al., 2018; Wardle and Derakhshan, 2017). Trust accounted for a significant indirect statistical association, and AI literacy moderated the trust-verification relationship. These findings support the relevance of studying AI-specific literacy while remaining subject to the limitations of cross-sectional, self-reported evidence.

7 Theoretical implications

This study contributes to trust theory by expanding the scope to AI-generated social media content. Traditional trust models tend to be based on interpersonal, organizational, or ecommerce relationships (Mayer et al., 1995; McKnight et al., 2002). By contrast, the current model focuses on trust in content whose producer might be non-human, hybrid, anonymous or hard to identify. Theoretically, this change is significant, as trust is no longer only placed in a person or an institution, but in an informational object, which is influenced by algorithmic processes.

The findings are compatible with information-processing accounts in which trust serves as a credibility shortcut and verification intention reflects more systematic evaluation (Kahneman, 2011; Petty and Cacioppo, 1986). However, this is not the only plausible interpretation. One might expect that individuals who already believe that AI-generated content is trustworthy would rationalize that they don't need to check it further, following a cognitive-dissonance explanation. Other factors that may lead to trust and decreased verification intention are effort avoidance, motivated reasoning, prior familiarity, and the use of the platform out of habit. These processes are not measured directly and so the data cannot differentiate among these competing explanations.

The study also contributes to the field of AI-literacy scholarship by explicitly linking AI-literacy to intention to verify as well as moderating the relationship between trust and verification. The direct term reflects differences in overall reported verification tendency, and the interaction term reflects whether there is a difference between the relationship between trust and verification across literacy levels. Rather, treating literacy as a mediator would mean that exposure or trust would influence literacy, which is a temporal process that this cross-sectional design cannot prove.

8 Practical implications

The results indicate the possibility of using clear labels on AI-disclosures, information about the origin of the content, source indicators, and timely reminders to verify the content as supports for user evaluation on platforms. Implementation is not easy: users could ignore or misinterpret labels, repeated warnings can lead to warning fatigue or habituation, and general warnings can give false reassurance, false distrust or stigma to legitimate AI-assisted content. Provenance systems can also be expensive, hard to find, or be deleted if the material is reposted. Any intervention should thus be usability-tested, accessible, proportionate to the level of risk and assessed for unintended consequences, not assumed.

The results would inform the incorporation of AI literacy into media literacy and digital communication education for educators and universities. Students are expected to learn how to evaluate AI-generated text, be aware of the limitations of AI, ask questions and verify facts. Courses may involve practical activities that raise ethical issues about synthetic media, try to identify AI-generated content, check facts, compare sources, etc. These courses need to be taught by trained instructors, have readily available materials and repeated practice activities; otherwise, the disparity in institutional resources may exacerbate existing literacy differences. The model recommends that, apart from platform accountability, there should also be a focus on capacity building of public sector in the regulation, as this is a key element of the model. Collectively, disclosure requirements, content sourcing policies, assistance for fact-checking entities and national AI literacy initiatives can help bolster the public's resistance to misleading AI-generated content.

The study shows that verification is a common digital action of the users and is of importance. Users are not to assume that the content is polished or realistic and should not be relied upon as accurate. They should check to confirm that information is accurate from a first-hand source, check information from multiple sources, watch for disclosure of AI by the source, and refrain from disseminating information that may affect the faith or decision-making process of others.

9 Limitations and future research

First, the cross-sectional design prevents causal inference and does not establish the temporal ordering of exposure, trust, literacy, and verification intention. The reported direct, indirect, and interaction estimates should therefore be interpreted as statistical associations. Longitudinal or experimental studies are needed to examine direction, change over time, and alternative causal explanations.

Second, all focal constructs were measured through self-report, and verification intention is socially desirable. Standardized, behavior-specific items and documented informed consent do not remove the possibility of socially desirable responding or common-method variance. No social-desirability scale, marker-variable test, or behavioral verification task is documented. Future studies should incorporate observed fact-checking, source inspection, click-tracking, or experimental sharing decisions, together with stronger procedural controls.

Third, the study relied on nonprobability convenience recruitment within a predominantly university-centered Jordanian setting, with 74.8% of respondents in the bachelor's-level academic category. Regional coverage does not establish national representativeness, and the lack of a documented sampling denominator prevents calculation of a response rate. Generalization to non-students, older adults, journalists, or other cultural contexts is therefore limited. Future research should use probability-based or comparative sampling where feasible.

A further transparency limitation is that the original response screening log was not retained. Consequently, the initial submission total, criterion-specific exclusions, duplicate-identification method, completion-time cutoff, and straight-line-response threshold cannot be independently audited. Future studies should preserve a dated screening flow and prespecified exclusion rules.

Fourth, the deliberately parsimonious framework does not include general media awareness, information overload, AI anxiety, perceived risk, platform trust, prior misinformation experience, or perceived source credibility. These constructs are not inherently redundant and may represent omitted alternative explanations or confounders. Future research should test their incremental contribution and compare different types of AI-generated text, images, video, voice synthesis, and deepfakes. In addition, the absence of a documented expert review or formal pilot limits evaluation of the original item-adaptation process.

10 Conclusion

Within a university-centered convenience sample in Jordan, greater reported exposure to AI-generated social media content was associated with lower trust in such content and stronger self-reported verification intention. Trust accounted for a significant indirect statistical association between exposure and verification intention, and AI literacy was both positively associated with verification intention and associated with a weaker negative trust-verification relationship. These findings provide context-specific evidence that exposure, content-specific trust, and AI literacy should be considered together when studying the evaluation of synthetic social media content. However, the cross-sectional design does not establish causality, self-reported intentions may differ from actual verification behavior and may reflect social desirability, and the predominantly university-centered convenience sample limits generalizability beyond the observed Jordanian higher-education context. Recommendations for AI-literacy education and platform transparency should therefore be treated as context-sensitive proposals requiring further behavioral, longitudinal, and experimental evaluation.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fhumd.2026.1927700/full#supplementary-material

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Sunghyun Chung, Texas A&M University, United States

Reviewed by: Abdul Khoso, Yangzhou University, China

Jauhar Wahyuni, Universitas Negeri Surabaya, Indonesia