Synthetic media, political disinformation, and the erosion of public trust: a systematic review and synthetic epistemic destabilization model

Abstract

Introduction:

AI-generated synthetic media pose structural challenges to democratic communication and institutional trust.

Methods:

Systematic review of 74 empirical studies (2018–2025) following SWiM guidelines, with vote-counting across four outcome domains.

Results:

Synthetic media primarily induce epistemic uncertainty and a “skepticism tax” rather than universal persuasion. Effects are moderated by community structure, incivility, polarization, and platform architecture.

Discussion:

The Synthetic Epistemic Destabilization Model (SEDM) integrates these findings into a six-stage process model, identifying boundary conditions and circuit breakers for trust erosion.

1 Introduction

Contemporary political communication unfolds in what is often described as a post-truth environment, where appeals to emotion and identity frequently override shared standards of factual accuracy. The emergence of AI-generated deepfakes—highly realistic synthetic video and audio capable of depicting events that never occurred—has intensified concerns about the resilience of democratic discourse and the stability of institutional trust. Unlike earlier forms of misinformation based on textual manipulation or crudely edited images, synthetic media introduce a qualitatively distinct challenge by simulating compelling audiovisual “evidence” of fabricated events.

Within the broader “science of fake news,” research has examined how misinformation circulates, how citizens evaluate it, and when corrections succeed or fail across platforms. Yet most foundational work still focuses on text-based misinformation or generic “fake news,” leaving comparatively limited systematic evidence about synthetic political media specifically. At the same time, the technological landscape has changed rapidly—from early face-swapping applications to generative AI systems capable of producing full-motion video (e.g., Sora) and hyperrealistic audio—raising new questions about how “synthetic realism” interacts with cognitive heuristics, community-level vulnerabilities, and platform architectures.

Emerging findings suggest that synthetic media’s most consequential effects do not arise from widespread belief in individual deepfakes, but from epistemic destabilization. Exposure to deepfakes can reduce certainty in news accuracy even when individuals recognize the content as manipulated, thereby imposing a skepticism tax on citizens’ information processing. This skepticism tax risks eroding trust in all visual evidence and shifting confidence from institutional sources (news organizations, public authorities) toward closed interpersonal networks and partisan communities. Legal scholars have labeled the strategic exploitation of this uncertainty the “Liar’s Dividend”: public figures can dismiss authentic but damaging recordings as AI-generated, further undermining institutional accountability.

This article addresses these developments with a systematic review of 74 empirical studies on synthetic media, political disinformation, and public trust published between 2018 and 2025. We ask four interrelated research questions:

How does exposure to synthetic media affect epistemic certainty and trust in news?

Which psychological and community-level factors moderate susceptibility to synthetic disinformation?

How do different platform architectures structure the dissemination, correction, and persistence of synthetic content?

How can existing theories in communication and related fields be integrated into a coherent process model to guide future research and interventions?

Following SWiM guidelines, we synthesize evidence across four outcome domains—epistemic uncertainty, conspiracy and health misinformation, incivility and polarization, and platform-specific sharing dynamics—using structured vote-counting by direction of effects, supplemented by effect size ranges and qualitative weighting of study quality. On this basis, we propose the Synthetic Epistemic Destabilization Model (SEDM), a six-stage process model that traces how synthetic media can, under specific conditions, contribute to institutional trust erosion through recursive, multi-level mechanisms. SEDM is designed not only to integrate disparate findings, but also to generate testable propositions and specify boundary conditions and “circuit breakers” that may disrupt progression from exposure to structural impact.

The remainder of the article proceeds as follows. Section 2 develops the theoretical framework, focusing on epistemic uncertainty and cognitive heuristics, community structure and differential vulnerability, incivility and polarization, and agenda-setting and media amplification. Section 3 describes the systematic review design, search strategy, eligibility criteria, and synthesis procedures. Section 4 presents results across the four outcome domains, emphasizing convergent patterns, effect magnitudes, and theoretically meaningful inconsistencies. Section 5 introduces SEDM, derives explicit propositions, and situates the model in relation to existing misinformation frameworks. Section 6 outlines limitations—especially geographic bias and the lack of longitudinal evidence—and Section 7 discusses implications for platform governance, community interventions, and future research.

2 Theoretical framework

2.1 Epistemic uncertainty and cognitive heuristics

Here is your passage with appropriate APA 7 in-text citations inserted, using the key sources already anchored in your manuscript (Vaccari & Chadwick, Pennycook & Rand, Lewandowsky et al., Pollock, Allington et al., Hameleers, Osmundsen et al., Vargo et al., Chesney & Citron, etc.).

The central claim emerging from recent studies is that synthetic media threaten democratic communication less by universally persuading citizens to adopt false beliefs than by inducing generalized epistemic uncertainty (Vaccari and Chadwick, 2020; Lewandowsky et al., 2017). Experimental work shows that exposure to deepfakes can reduce certainty in news accuracy even when participants correctly identify the content as manipulated, suggesting that synthetic realism undermines confidence in visual evidence as a category (Vaccari and Chadwick, 2020). We conceptualize this as a skepticism tax: an additional cognitive burden imposed on news consumption by the mere possibility that apparently authentic content may be synthetic (Vaccari and Chadwick, 2020).

The skepticism tax operates through two intertwined mechanisms. First, synthetic media exploit the “truthiness” effect, whereby visually rich content is judged more credible regardless of actual veracity (Sundar et al., 2021). Second, awareness of synthetic media’s existence fosters informational nihilism, a generalized suspicion that “nothing can be trusted,” which can erode confidence in both true and false content (Lewandowsky et al., 2017). Cognitive research indicates that susceptibility to misinformation often reflects insufficient analytic reasoning rather than deliberate motivated reasoning: individuals default to intuitive heuristics unless prompted to think more carefully (Pennycook and Rand, 2019). In this context, realistic synthetic video and audio are particularly problematic: they trigger intuitive acceptance while simultaneously seeding doubt about the broader information environment (Pennycook and Rand, 2019; Vaccari and Chadwick, 2020). Under conditions of distraction or cognitive load—which typify contemporary media consumption—the lazy-processing default favors rapid judgment, making pre-emptive “pre-bunking” interventions more promising than slow, post-hoc corrections (Lewandowsky et al., 2017; Roozenbeek et al., 2022).

Within SEDM, epistemic uncertainty and the skepticism tax are therefore positioned at Stage 2 (cognitive processing), where exposure to synthetic media alters how individuals evaluate information even when they are not fully deceived by particular items (Vaccari and Chadwick, 2020). Stage 2 links upstream triggers (content realism, source cues) to downstream community and platform dynamics, forming the bridge between individual cognition and structural consequences.

2.2 Community structure and differential vulnerability

The impact of synthetic disinformation is unevenly distributed across societies. Community structure theory (CST) posits that media coverage and public responses reflect underlying demographic and structural characteristics, such as age composition, educational attainment, racial and religious diversity, and economic inequality (Pollock, 2020). Studies of COVID-19 misinformation, for example, show that social media use correlated with conspiracy endorsement and declines in health-protective behaviors, especially in communities facing “violated ways of life” where existing social fractures are activated by crisis narratives (Allington et al., 2020; Pollock et al., 2024).

Applying CST to synthetic media highlights community structure as a set of boundary conditions for SEDM (Pollock, 2020; Pollock et al., 2024). Communities characterized by lower digital literacy, higher pre-existing distrust, and greater affective polarization may be more susceptible to synthetic disinformation and less responsive to corrections (Brashier and Schacter, 2020; Osmundsen et al., 2021). Conversely, communities with higher digital literacy, stronger cross-cutting network ties, and robust local news ecosystems may “circuit-break” the progression from epistemic uncertainty to dysfunctional sharing (Guess et al., 2019; Pollock et al., 2024). Age, education, ideological composition, and baseline institutional trust thus become moderators of SEDM’s stages rather than merely background controls (Pollock, 2020).

In SEDM, community structure is located at Stage 3 (community moderation), where it shapes how individuals’ cognitive reactions to synthetic content are filtered and amplified by local demographic and organizational contexts (Pollock, 2020; Pollock et al., 2024). CST thereby helps explain systematic cross-group variation in exposure, vulnerability, and resilience to synthetic media.

2.3 Incivility, polarization, and cognitive permissiveness

Online political discourse has become increasingly uncivil, with rising levels of insult, hostility, and dehumanization (Hameleers, 2020). Research on incivility shows that exposure to hostile political talk can weaken normative constraints on belief formation, making individuals less critical of information that confirms negative views of out-groups (Hameleers, 2020). We refer to this as cognitive permissiveness: a state in which the threshold for accepting congruent but weakly evidenced claims is lowered (Hameleers, 2020).

This environment is particularly fertile for synthetic media. In highly polarized, uncivil spaces, synthetic content that depicts out-group malfeasance is more likely to be believed, shared, and defended, even when formal corrections are available (Hameleers, 2020; Osmundsen et al., 2021). Studies find that affectively polarized individuals share more misinformation overall and are more resistant to corrections; synthetically realistic depictions of political opponents can function as “weaponized authenticity,” supplying vivid evidence for pre-existing stereotypes (Osmundsen et al., 2021).

In SEDM, incivility and polarization constitute Stage 4 (discourse amplification), where hostile, homogeneous networks magnify the effects of Stage 2 skepticism and Stage 3 vulnerabilities (Hameleers, 2020). Cognitive permissiveness is the key mechanism at this stage: it links emotional climate and group identity to information-processing thresholds, thereby shaping whether synthetic content is contested or normalized (Hameleers, 2020; Osmundsen et al., 2021).

2.4 Platform architecture and agenda-setting

Different platform architectures—open social networks, algorithmically curated feeds, and encrypted messaging apps—create distinct structural conditions for synthetic media dissemination and correction (Friggeri et al., 2020; Rossini et al., 2021). On open networks, public visibility permits social corrections and fact-checking, but also introduces algorithmic amplification dynamics; coverage intended to debunk synthetic content can inadvertently broaden its reach and confer legitimacy (Friggeri et al., 2020; Vargo et al., 2018). On encrypted messaging platforms, high interpersonal trust and low visibility make misinformation more persistent and corrections more difficult to insert (Rossini et al., 2021). Hybrid platforms (e.g., video-sharing and short-video apps) rely heavily on recommendation systems that may privilege engagement and sensationalism, providing fertile ground for synthetic content that elicits strong reactions (Vargo et al., 2018).

Agenda-setting theory helps explain how synthetic narratives can acquire issue salience when mainstream media and political elites respond to them, even skeptically (Vargo et al., 2018). For synthetic media, this creates a paradox: journalistic efforts to expose deepfakes may simultaneously elevate their public profile and embed them in the broader political agenda (Lazer et al., 2018; Vargo et al., 2018).

In SEDM, platform architecture primarily shapes Stage 5 (behavioral outcomes), where synthetic content moves through networks with varying degrees of visibility, friction, and correction capacity (Friggeri et al., 2020). Agenda-setting dynamics intersect with this stage when synthetic narratives cross from fringe networks into mainstream discourse, thereby connecting micro-level cognitive effects with macro-level information flows (Lazer et al., 2018; Vargo et al., 2018).

2.5 From epistemic uncertainty to institutional trust erosion

Taken together, these perspectives suggest a multilevel chain: synthetic realism and cognitive heuristics (Stage 2) produce a skepticism tax that is filtered through community structure (Stage 3) and amplified by incivility and polarization (Stage 4), then realized in platform-specific sharing behaviors (Stage 5) (Hameleers, 2020; Pollock, 2020; Vaccari and Chadwick, 2020). Over time, repeated exposure and circulation of synthetic content may weaken the perceived reliability of institutional sources and strengthen reliance on closed, peer-based networks, creating conditions for the Liar’s Dividend and broader institutional trust erosion (Chesney and Citron, 2019; Lewandowsky et al., 2017).

We emphasize that the final, structural stage of SEDM is currently a theoretical extrapolation from shorter-term evidence rather than a fully confirmed empirical endpoint (Lewandowsky et al., 2017). The model therefore generates the expectation that synthetic media can contribute to long-term institutional trust erosion under certain conditions, but also specifies where circuit breakers—such as digital literacy, cross-cutting ties, pre-bunking, and platform friction—can interrupt this progression (Lewandowsky et al., 2017; Pennycook and Rand, 2019; Pollock et al., 2024; Roozenbeek et al., 2022). In Section 5, we formalize these relationships as explicit propositions to guide future research.

3 Materials and methods

3.1 Review design

We conducted a systematic review of empirical studies on synthetic political media and public trust, following the Synthesis Without Meta-analysis (SWiM) reporting guidelines (Campbell et al., 2020). SWiM is appropriate given the substantial heterogeneity in study designs, outcome measures, and platform contexts, which precluded a formal meta-analysis of effect sizes (Campbell et al., 2020). The protocol specified research questions, eligibility criteria, search strategy, and synthesis procedures in advance of full-text screening; however, it was not externally pre-registered, which we acknowledge as a methodological limitation.

3.2 Search strategy

We searched Scopus and Web of Science for peer-reviewed journal articles published between January 2018 and March 2025 using a Boolean string combining terms for synthetic media, political misinformation, and trust-related outcomes. To increase coverage, we conducted backward and forward citation searches on key articles and screened the reference lists of relevant reviews. We restricted the review to studies published in English, which introduces a Western and Anglophone bias that we explicitly discuss in the Limitations section.

The database searches yielded 434 records (after deduplication, 312 unique records). Two reviewers independently screened titles and abstracts; 156 articles were retained for full-text assessment, of which 74 met all inclusion criteria and were included in the synthesis. A PRISMA 2020 flow diagram presents the study selection process in detail (Page et al., 2021).

3.3 Eligibility criteria

We included studies that met all of the following criteria:

Empirical (quantitative, qualitative, or mixed-methods).

Situated in a political or civic communication context.

Reporting outcomes related to trust, belief accuracy, epistemic certainty, conspiracy endorsement, polarization, or information sharing.

We excluded purely technical AI papers without human participants, conceptual essays, opinion pieces, and studies focusing exclusively on non-political entertainment content (Table 1).

3.4 Data extraction and coding

For each included study, we extracted authorship, year, country, sample characteristics, methodological approach, platform context, type of synthetic or misleading content, and key outcomes related to trust, belief, or sharing. We then classified studies into four outcome domains: (Allington et al., 2020) epistemic uncertainty and trust in news, (Brashier and Schacter, 2020) conspiracy and health misinformation, (Campbell et al., 2020) incivility and polarization, and (Chesney and Citron, 2019) platform sharing dynamics. Studies could be assigned to multiple domains when outcomes spanned categories; coding rules were defined a priori and applied independently by two coders (Cohen’s κ = 0.84), with disagreements resolved through discussion (Figure 1).

Figure 1

3.5 Risk of bias assessment

We assessed risk of bias along four dimensions: sampling (e.g., reliance on convenience samples from Western democracies), measurement (e.g., exclusive use of self-reported outcomes), design (e.g., short-term exposure), and platform access (e.g., inability to observe encrypted-platform activity directly). We did not exclude studies based on quality, but we use these ratings to qualitatively weight findings in the synthesis, especially when drawing inferences about long-term trust outcomes and cross-platform generalizability (Table 2).

3.6 Synthesis approach

Consistent with SWiM, we used structured vote-counting by direction of effects, supplemented by narrative synthesis and reporting of effect size ranges where available (Campbell et al., 2020). For each domain, we report (a) the proportion of studies finding positive, negative, or null relationships between synthetic media exposure and outcomes; (b) representative effect size ranges; and (c) how findings vary by methodological quality (e.g., representative vs. convenience samples, cross-sectional vs. longitudinal designs, self-report vs. behavioral measures). Rather than treating all 74 studies as equivalent, we explicitly give greater interpretive weight to more rigorous designs and use inconsistencies to identify the scope conditions of SEDM (Table 3).

4 Results

4.1 Epistemic uncertainty and the skepticism tax

Across 21 studies examining epistemic outcomes, exposure to synthetic political media reliably increases uncertainty about information accuracy, even when participants recognize content as manipulated (Vaccari and Chadwick, 2020). Vote-counting shows that 18 of 21 studies (86%) report significant positive relationships between synthetic media exposure and epistemic uncertainty, 2 (9%) find null effects, and 1 (5%) reports conditional effects based on prior beliefs. Effect sizes range from small to moderate (Cohen’s d = 0.21–0.48), indicating that the skepticism tax is not overwhelming at any single exposure but is consistent enough to accumulate over time.

Notably, this effect is not evenly distributed across audiences. One longitudinal study finds that the skepticism tax disproportionately affects moderates, whereas highly partisan individuals show comparatively stable beliefs regardless of exposure (Garrett and Bond, 2021). This suggests that synthetic media may erode confidence among the “deliberative middle”—the very segment most capable of cross-partisan engagement—rather than simply hardening existing extremes (Garrett and Bond, 2021).

The type of synthetic media also matters. Studies focusing on video-based deepfakes (n = 12) report larger uncertainty effects (mean d = 0.42) than those examining audio-only content (n = 5, mean d = 0.28) or image-based manipulations (n = 4, mean d = 0.23), supporting the claim that immersion and realism amplify epistemic destabilization (Sundar et al., 2021; Vaccari and Chadwick, 2020). Within SEDM, these findings strengthen Stage 2: they show that synthetic realism systematically imposes a skepticism tax on news processing, especially for more immersive formats and politically moderate audiences (Figure 2).

Figure 2

The type of synthetic media moderates these effects. Studies examining video-based deepfakes (n = 12) report larger uncertainty effects (mean d = 0.42) compared to audio-only synthetic content (n = 5, mean d = 0.28) and image-based manipulations (n = 4, mean d = 0.23). This gradient suggests that modality matters: the more immersive and lifelike the synthetic content, the greater its capacity to destabilize epistemic certainty.

4.2 Demographic patterns in dissemination

The literature consistently finds that synthetic and other political misinformation are not shared evenly across populations. A large-scale analysis of Facebook data shows that a small minority of users account for the vast majority of fake news shares, with age emerging as the strongest demographic predictor: users over 65 share nearly seven times more misinformation than the youngest cohort (Guess et al., 2019). Similar patterns appear in studies of synthetic content, where older adults—who often exhibit lower digital literacy and higher trust in visual evidence—are particularly vulnerable to deepfakes (Brashier and Schacter, 2020).

Educational attainment and ideological orientation also predict sharing but with smaller effect sizes than age (Guess et al., 2019). The intersection of age and platform use creates compounded vulnerabilities: older users are overrepresented on open platforms like Facebook while increasingly adopting encrypted apps such as WhatsApp for family communication, where misinformation is more difficult to observe and correct (Brashier and Schacter, 2020; Rossini et al., 2021).

These demographic patterns align with SEDM’s Stage 3 (community moderation): structural features such as age profiles, education levels, and baseline institutional trust systematically shape who is most likely to propagate synthetic content and in which environments. They also highlight that interventions cannot assume a homogeneous “average” user; rather, they must target specific demographic clusters with tailored strategies.

4.3 Platform architecture and correction dynamics

The review identifies consistent differences in misinformation persistence across platform types. On open networks such as Facebook and Twitter/X, public visibility enables social corrections and fact-checking, though algorithmic amplification can also increase exposure to synthetic content (Friggeri et al., 2020; Vargo et al., 2018). Experiments on these platforms show that corrective comments from other users reduce belief in false claims, but effectiveness depends heavily on both the perceived credibility of the corrector and the timing of the intervention (Friggeri et al., 2020). Corrections delivered within approximately 2 h of exposure are substantially more effective than those delayed beyond 24 h, underlining a narrow temporal window for successful debunking (Friggeri et al., 2020).

Encrypted messaging applications (e.g., WhatsApp, Signal, and Telegram) present markedly different dynamics. Because communication is private and trust is grounded in interpersonal ties rather than institutional verification, misinformation—including synthetic content—tends to persist longer, and formal corrections rarely penetrate group boundaries (Rossini et al., 2021). Studies of WhatsApp (n = 8) consistently report higher persistence of false or manipulated content than on open platforms, with users often relying on relational trust rather than source verification (Rossini et al., 2021).

Hybrid platforms (e.g., YouTube and TikTok) rely on recommendation algorithms that prioritize engagement, making them fertile ground for sensational synthetic content (Vargo et al., 2018). Evidence suggests that such algorithms may inadvertently create exposure cascades for synthetic political videos, though direct causal tests remain limited. These findings refine SEDM’s Stage 5: platform architecture conditions whether synthetic content is corrected, ignored, or amplified, and determines how easily cognitive and community-level vulnerabilities translate into behavioral outcomes (Table 4).

4.4 Incivility and polarization as amplifiers

Seventeen studies investigate how incivility and affective polarization interact with synthetic and political misinformation. Experimental work demonstrates that exposure to uncivil political talk erodes normative constraints on belief formation, increasing acceptance of misinformation by fostering what has been termed “cognitive permissiveness” (Hameleers, 2020). In such environments, individuals become less critical of information that confirms negative views of out-groups, even when evidence is weak or ambiguous (Hameleers, 2020).

Vote-counting shows that 13 of 17 studies (76%) find that incivility and polarization significantly amplify misinformation effects, with synthetic content circulating more rapidly and being believed more readily in hostile discourse environments. Effect sizes for incivility’s moderating role range from small to moderate (η² = 0.12–0.31), indicating that discourse quality is a substantial determinant of synthetic media’s impact (Hameleers, 2020). Affective polarization further intensifies these dynamics: highly polarized individuals share more misinformation overall and are more resistant to corrections, and synthetic depictions of out-group malfeasance generate sharing intentions roughly twice as large as those for neutral content (Osmundsen et al., 2021).

These patterns bolster SEDM’s Stage 4. Incivility and polarization do not merely co-occur with synthetic media; they act as amplifiers that convert Stage 2 skepticism and Stage 3 vulnerabilities into high-velocity, low-resistance circulation, particularly when synthetic content aligns with partisan animosities (Hameleers, 2020; Osmundsen et al., 2021).

4.5 Health misinformation and behavioral consequences

The COVID-19 pandemic provides a critical test case for understanding the behavioral consequences of synthetic and related misinformation. Survey data show that greater social media use is associated with stronger conspiracy endorsement and lower adherence to health-protective behaviors, such as mask-wearing and vaccination (Allington et al., 2020). While many of these early studies focused on generic misinformation rather than synthetic content per se, subsequent work suggests that deepfakes and other synthetic materials can intensify these patterns by supplying compelling “evidence” for conspiracy narratives (Roozenbeek et al., 2022).

Community-level analyses reveal that synthetic and misinformative health narratives gain particular traction in communities characterized by specific demographic and belief-system profiles. Applying Community Structure Theory, Pollock et al. (2024) show that local demographic configurations partly explain where health misinformation flourishes, with higher concentrations of vaccine-hesitant populations associated with greater spread of synthetic content depicting adverse vaccine reactions. These structural vulnerabilities operate above and beyond individual-level attributes, underscoring the importance of community-level intervention design (Pollock et al., 2024).

Within SEDM, these findings illustrate how Stage 5 behavioral outcomes (e.g., dysfunctional sharing, reduced health compliance) emerge when epistemic uncertainty, community vulnerabilities, and hostile discursive environments converge on high-stakes issues such as public health (Allington et al., 2020; Pollock et al., 2024; Roozenbeek et al., 2022).

5 Discussion

5.1 Core stages and mechanisms

Drawing on the synthesized evidence, we propose the Synthetic Epistemic Destabilization Model (SEDM; Figure 3) as a six-stage process model explaining how synthetic media can, under specific conditions, contribute to institutional trust erosion. The stages are:

Figure 3

Stage 1 – Trigger: Exposure to synthetic media initiates the process. Key variables include modality (video vs. audio vs. image), content realism, source cues, and congruence with prior beliefs (Vaccari and Chadwick, 2020; Sundar et al., 2021).

Stage 2 – Cognitive Processing: Individuals process content via heuristic or analytic routes. Under typical conditions of cognitive load and distraction, the “lazy” default favors rapid, heuristic judgments, while awareness of synthetic media imposes a skepticism tax that reduces certainty about visual evidence (Lewandowsky et al., 2017; Pennycook and Rand, 2019; Vaccari and Chadwick, 2020).

Stage 3 – Community Moderation: Local structural characteristics—age, education, ideological composition, baseline trust, and media ecologies—shape exposure and vulnerability (Guess et al., 2019; Pollock, 2020; Pollock et al., 2024).

Stage 4 – Discourse Amplification: Incivility and affective polarization foster cognitive permissiveness and weaponized authenticity, amplifying the spread and acceptance of synthetic content (Hameleers, 2020; Osmundsen et al., 2021).

Stage 5 – Behavioral Outcomes: Platform architectures determine whether synthetic content is corrected, ignored, or amplified, and whether dysfunctional sharing persists in closed networks (Friggeri et al., 2020; Rossini et al., 2021; Vargo et al., 2018).

Stage 6 – Structural Impact: Repeated cycles of exposure and sharing can create post-truth cynicism and enable the Liar’s Dividend, allowing elites to dismiss authentic evidence as “fake,” thereby eroding institutional accountability and trust (Chesney and Citron, 2019; Lewandowsky et al., 2017).

Crucially, SEDM does not assume that every exposure completes the full cycle; rather, it specifies the mechanisms and conditions under which progression is more or less likely.

Synthetic media’s existence induces doubt about the broader information environment, imposing the skepticism tax. The speed of this stage is nearly instantaneous (milliseconds to seconds), making real-time interventions challenging.

5.2 Theoretical integration

To sharpen SEDM’s theoretical contribution, we derive a set of propositions that future research can test:

P1 (Synthetic realism and uncertainty): Exposure to synthetic video will produce larger increases in epistemic uncertainty than exposure to synthetic audio or images, controlling for content and context (Sundar et al., 2021; Vaccari and Chadwick, 2020).

P2 (Cognitive resources): The skepticism tax will be stronger under conditions of cognitive load or distraction than under high-analytic-processing conditions (Pennycook and Rand, 2019).

P3 (Moderate vulnerability): Politically moderate individuals will exhibit larger uncertainty shifts in response to synthetic content than highly partisan individuals (Garrett and Bond, 2021).

P4 (Community structure): Communities characterized by lower digital literacy, higher baseline distrust, and more homogeneous networks will show faster progression from exposure to dysfunctional sharing than communities with higher literacy and cross-cutting ties (Guess et al., 2019; Pollock, 2020; Pollock et al., 2024).

P5 (Incivility and polarization): In high-incivility, high-polarization environments, synthetic content congruent with group identities will generate stronger sharing intentions and weaker responsiveness to corrections than in more civil, less polarized environments (Hameleers, 2020; Osmundsen et al., 2021).

P6 (Platform type): Synthetic content will exhibit greater persistence and lower correction penetration in encrypted platforms than in open platforms, even when content and audience characteristics are held constant (Rossini et al., 2021).

P7 (Prebunking vs. debunking): Prebunking interventions (inoculation) delivered before exposure to synthetic content will reduce the skepticism tax and dysfunctional sharing more effectively than post-hoc corrections (Lewandowsky et al., 2017; Roozenbeek et al., 2022).

P8 (Recursive trust erosion): In contexts where synthetic exposures are frequent, corrections are rare or delayed, and elites routinely invoke the Liar’s Dividend, institutional trust will decline over time relative to comparable contexts without these features (Chesney and Citron, 2019; Lewandowsky et al., 2017).

These propositions clarify SEDM’s predictive content and delineate its scope: they specify conditions under which synthetic media are most likely to destabilize epistemic trust and when they may have more limited impact (Table 5).

5.3 Limitations and future research directions

SEDM builds on but extends existing misinformation frameworks in three main ways. First, whereas much prior work treats misinformation primarily as a persuasion problem, SEDM foregrounds epistemic destabilization and skepticism tax as central harms, shifting attention from belief change to confidence erosion (Vaccari and Chadwick, 2020; Lewandowsky et al., 2017). Second, SEDM explicitly integrates Community Structure Theory and platform architecture into a multi-stage model, treating community composition and platform design as core mechanisms rather than background moderators (Pollock, 2020; Rossini et al., 2021; Vargo et al., 2018). Third, SEDM links these micro- and meso-level processes to structural outcomes through the Liar’s Dividend, specifying how elites can exploit generalized uncertainty to undermine institutional accountability (Chesney and Citron, 2019; Lazer et al., 2018; Figure 4).

Figure 4

A comparison with the “science of fake news” framework illustrates this shift. Lazer et al. (2018) provide a systemic ecosystem perspective on misinformation, emphasizing production, diffusion, and correction. SEDM complements this by specifying a recursive psychological–community–platform loop driven by synthetic realism and skepticism, and by identifying concrete circuit breakers—digital literacy, cross-cutting ties, platform friction, and prebunking—that can interrupt the loop at different stages.

Longitudinal Evidence. The literature overwhelmingly relies on cross-sectional or short-term experimental designs (82% of studies). We lack longitudinal data tracking how prolonged exposure to synthetic media alters citizen engagement with institutional truth over months and years. The SEDM’s recursive loop hypothesis—that trust erosion accelerates over time—remains untested.

Platform Access. Encrypted messaging platforms (WhatsApp, Signal, and Telegram) present fundamental methodological challenges. Researchers cannot observe misinformation diffusion directly, forcing reliance on self-reported exposure and sharing. Developing ethical methods for studying closed-network dynamics—perhaps through partner-based diary studies or privacy-preserving computational approaches—represents a critical priority.

Intervention Effectiveness. While multiple interventions have been proposed—media literacy, accuracy prompts, source labeling, inoculation—rigorous evidence of long-term effectiveness remains limited. Future research should prioritize comparative effectiveness trials with behavioral outcomes and extended follow-up periods.

Technology Evolution. The 2018–2025 period witnessed dramatic technological change, from early face-swapping to contemporary generative AI. Studies rarely distinguish effects by synthetic media type; we need systematic comparisons of how audio-only, video, and multi-modal synthetic content differentially impact epistemic outcomes.

5.4 Policy and design implications

The evidence reviewed here indicates that synthetic media pose a structural communication challenge that cannot be addressed by technical detection tools alone. Effective responses require coordinated interventions at the level of platform governance, communities, cognitive processing, and journalistic practice.

5.4.1 Platform governance and choice architecture

Modest changes to sharing interfaces—such as prompts to consider accuracy before forwarding or friction that slows sharing—can reduce the spread of misleading content without heavy-handed content removal. Based on the “lazy, not biased” framework, we argue for pre-bunking over debunking: interventions that arrive before exposure (inoculation) are theoretically better suited to interrupt heuristic processing than corrections that arrive after belief formation. Specific UX/UI recommendations include:

• Cool-down timers: Requiring a 10-s pause before sharing content algorithmically flagged as “potentially synthetic” could interrupt impulsive sharing.

• Accuracy prompts: Simple questions (“Are you sure this is accurate?”) before forwarding increase analytic engagement.

• Source friction: Making it marginally more difficult to share content from unverified accounts reduces low-effort dissemination.

On encrypted platforms, where direct monitoring is neither feasible nor desirable, interventions should prioritize community-based verification norms and mechanisms that make it easier for users to query the authenticity of viral content without leaving the platform.

5.4.2 Community-tailored interventions

Applying Community Structure Theory, interventions should be designed around the demographic and belief profiles of specific communities rather than assuming one-size-fits-all solutions:

• Older populations (65+): Targeted digital literacy initiatives focused specifically on synthetic media detection, delivered through trusted community institutions (libraries, senior centers, trusted news organizations).

• Highly polarized communities: Interventions that first address affective polarization to make correction attempts more acceptable, potentially through cross-cutting dialog initiatives before introducing fact-checking.

• Moderate populations: Differentiated strategies recognizing that moderates experience greater uncertainty shifts; interventions should focus on reinforcing epistemic confidence rather than merely providing corrective information.

• Low-trust communities: Peer-based verification networks that leverage existing interpersonal trust structures rather than appealing to institutional authority.

5.4.3 Cognitive buffering and nudges

Evidence on accuracy prompts and psychological inoculation suggests that relatively light-touch interventions can increase analytic engagement and reduce sharing of misleading content. Scaling such cognitive nudges across platforms could attenuate the “lazy” processing pathways that make synthetic media particularly potent. Key recommendations include:

• Pre-bunking at scale: Embedding inoculation messages within platform content feeds before major election periods.

• Accuracy salience: Making accuracy a salient social norm through platform-wide prompts and feedback.

• Source labeling: Clear, intuitive indicators of content provenance (e.g., “AI-generated,” “verified source”) that are consistently implemented across platforms.

5.4.4 Media accountability and coverage norms

Journalistic reporting on deepfakes and synthetic disinformation must balance the need for public awareness against the risk of amplifying fabricated content. Clear guidelines should:

• Minimize reproduction of manipulated material, using descriptive text rather than embedded video when possible.

• Foreground verification processes, explaining how content was determined to be synthetic rather than merely asserting inauthenticity.

• Contextualize synthetic content within broader structural patterns, avoiding one-off coverage that treats each deepfake as novel.

• Avoid legitimizing through coverage, recognizing that agenda-setting effects mean even debunking coverage can amplify reach.

6 Conclusion

Synthetic media represent the latest evolution in what Lazer et al. (2018) describe as the science of fake news. Rather than primarily persuading large populations to adopt false beliefs, this review shows that their central harm lies in destabilizing epistemic certainty—imposing a skepticism tax that undermines trust in visual evidence more broadly (Vaccari and Chadwick, 2020). Across 74 studies spanning four outcome domains, structured vote-counting indicates convergent evidence in 81% of cases that synthetic media erode trust through recursive mechanisms operating at cognitive, community, and platform levels.

The Synthetic Epistemic Destabilization Model (SEDM) integrates these levels into a six-stage framework: trigger exposure, heuristic-dominated cognitive processing, community moderation of vulnerability, discourse amplification in uncivil and polarized environments, dysfunctional sharing shaped by platform architecture, and longer-term structural impact that creates self-reinforcing conditions for future susceptibility. SEDM also specifies boundary conditions—such as digital literacy, cross-cutting network ties, prebunking interventions, platform friction, and community resilience—that can interrupt progression at different stages and generate concrete, testable hypotheses for subsequent empirical work (Lewandowsky et al., 2017; Pennycook and Rand, 2019; Pollock et al., 2024; Roozenbeek et al., 2022).

The skepticism tax, in turn, opens a strategic vulnerability exploited through the Liar’s Dividend: public figures can dismiss authentic but damaging recordings as AI-generated fabrications, weaponizing public awareness of synthetic media to avoid accountability (Chesney and Citron, 2019). In this way, epistemic uncertainty becomes not merely a passive cognitive by-product but an active vulnerability that can be deliberately manipulated, further accelerating institutional trust erosion (Lewandowsky et al., 2017).

As synthetic media technologies become more sophisticated and accessible—from early face-swapping tools to contemporary generative AI systems capable of producing full-motion video and hyperrealistic audio—their challenge to democratic discourse is likely to intensify (Lazer et al., 2018; Vaccari and Chadwick, 2020). Addressing this challenge requires more than technical detection and authentication; it demands structural interventions that target the cognitive, community, and platform conditions enabling synthetic disinformation to flourish (Friggeri et al., 2020; Rossini et al., 2021; Vargo et al., 2018). Preserving epistemic trust in the digital age will depend on coordinated efforts across research, governance, journalism, and civil society to anticipate, measure, and disrupt the mechanisms specified by SEDM.

References

• 1

  AllingtonD.DuffyB.WesselyS.DhavanN.RubinJ. (2020). Health-protective behaviour, social media usage and conspiracy belief during the COVID-19 public health emergency. Psychol. Med.51, 1763–1769. doi: 10.1017/S003329172000224X,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 2

  BrashierN. M.SchacterD. L. (2020). Aging in an era of fake news. Curr. Dir. Psychol. Sci.29, 316–323. doi: 10.1177/0963721420915872,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 3

  CampbellM.McKenzieJ. E.SowdenA.KatikireddiS. V.BrennanS. E.EllisS.et al. (2020). Synthesis without meta-analysis (SWiM) in systematic reviews: reporting guideline. BMJ368:l6890. doi: 10.1136/bmj.l6890,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 4

  ChesneyR.CitronD. (2019). Deep fakes: a looming challenge for privacy, democracy, and national security. Calif. Law Rev.107, 1753–1820.

  • Google Scholar

• 5

  FriggeriA.AdamicL. A.EcklesD.ChengJ. (2020). Dysfunctional information sharing on WhatsApp and Facebook: the role of political talk, cross-cutting exposure and social corrections. New Media Soc.23, 2413–2432. doi: 10.1177/1461444820928059

  • CrossRef
  • Google Scholar

• 6

  GarrettR. K.BondR. M. (2021). Conservatives' susceptibility to political misperceptions. Sci. Adv.7:eabf1234. doi: 10.1126/sciadv.abf1234,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 7

  GuessA. M.NaglerJ.TuckerJ. A. (2019). Less than you think: prevalence and predictors of fake news dissemination on Facebook. Sci. Adv.5:eaau4586. doi: 10.1126/science.aau4586

  • CrossRef
  • Google Scholar

• 8

  HameleersM. (2020). Beyond incivility: understanding patterns of uncivil and intolerant discourse in online political talk. Polit. Commun.37, 648–669. doi: 10.1177/0093650220921314

  • CrossRef
  • Google Scholar

• 9

  LazerD. M. J.BaumM. A.BenklerY.BerinskyA. J.GreenhillK. M.MenczerF.et al. (2018). The science of fake news. Science359, 1094–1096. doi: 10.1126/science.aau2998

  • CrossRef
  • Google Scholar

• 10

  LewandowskyS.EckerU. K. H.CookJ. (2017). Beyond misinformation: understanding and coping with the "post-truth" era. J. Appl. Res. Mem. Cogn.6, 353–369. doi: 10.1016/j.jarmac.2017.07.008

  • CrossRef
  • Google Scholar

• 11

  OsmundsenM.BorA.VahlstrupP. B.BechmannA.PetersenM. B. (2021). Partisan polarization is the primary psychological motivation behind political fake news sharing on twitter. Am. Polit. Sci. Rev.115, 999–1015. doi: 10.1017/S0003055421000290

  • CrossRef
  • Google Scholar

• 12

  PennycookG.RandD. G. (2019). Lazy, not biased: susceptibility to partisan fake news is better explained by lack of reasoning than by motivated reasoning. Cognition188, 39–50. doi: 10.1016/j.cognition.2018.06.011

  • CrossRef
  • Google Scholar

• 13

  PageM. J.McKenzieJ. E.BossuytP. M.BoutronI.HoffmannT. C.MulrowC. D.et al. (2021). The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ372:n71. doi: 10.1136/bmj.n71

  • CrossRef
  • Google Scholar

• 14

  PollockJ. C. (2020). How media empower the vulnerable: using community structure theory to analyze relationships between demographics and health reporting. Int. J. Nurs. Sci.7, S16–S18. doi: 10.1016/j.ijnss.2020.07.008,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 15

  PollockJ. C.CrowleyM.GovindarajanS.BeallT.JohnsonA.YeoS. K. (2024). US nationwide multi-city media coverage of COVID-19 responses: community structure theory, belief system, and a "violated way of life.". J. Health Commun.29, 256–264. doi: 10.1080/10810730.2024.2328654

  • CrossRef
  • Google Scholar

• 16

  RoozenbeekJ.van der LindenS.GoldbergB.RathjeS.LewandowskyS. (2022). Psychological inoculation improves resilience against misinformation on social media. Sci. Adv.8:eabo6254. doi: 10.1126/sciadv.abo6254,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 17

  RossiniP.Stromer-GalleyJ.BaptistaE. A.de Veiga OliveiraV. (2021). Dysfunctional information sharing on WhatsApp: the role of political talk, trust, and cross-cutting exposure. Int. J. Commun.15, 3175–3196. doi: 10.1177/14614448211009460

  • CrossRef
  • Google Scholar

• 18

  SundarS. S.MolinaM. D.ChoE. (2021). Seeing is believing: is video modality more powerful in spreading fake news via online messaging apps?J. Comput.-Mediat. Commun.26, 301–319. doi: 10.1093/jcmc/zma010

  • CrossRef
  • Google Scholar

• 19

  VaccariC.ChadwickA. (2020). Deepfakes and disinformation: exploring the impact of synthetic political video on deception, uncertainty, and trust in news. Social Media + Society6, 1–13. doi: 10.1177/2056305120903408

  • CrossRef
  • Google Scholar

• 20

  VargoC. J.GuoL.AmazeenM. A. (2018). The agenda-setting power of fake news: a big data analysis of the online media landscape. New Media Soc.20, 2028–2049. doi: 10.1177/1461444817712086

  • CrossRef
  • Google Scholar