When relief valves leak: directed digital dissidence, curated networks, and critical information exposure across media systems

1 Introduction

In semi-authoritarian regimes, citizens navigate an information environment still dominated by state-controlled media, yet increasingly supplemented by digital platforms. Many now construct their own curated information networks by following preferred sources and like-minded peers on social media. This personalization of news exposure fosters what scholars have called “echo chambers” or “filter bubbles,” where users are shielded from dissenting perspectives and mainly encounter congenial viewpoints (Bakshy et al., 2015). Prior research on selective exposure and avoidance shows that individuals tend to seek information that aligns with their preexisting beliefs and avoid conflicting content (Garrett, 2009). As a result, curated networks may limit exposure to government-critical content, unintentionally supporting regime narratives. Yet digital media also open new pathways for political expression. During the Arab Spring, for example, social media helped citizens coordinate protests and challenge state information monopolies (Howard and Hussain, 2011).

Authoritarian regimes, however, are not passive in the face of digital transformation. They have developed strategies to manage, redirect, and contain online discourse. Gainous et al. (2023) argue that scholars have focused extensively on how social media empowers dissidents, but paid less attention to how autocrats strategically manage digital spaces to maintain control. Their concept of Directed Digital Dissidence (DDD) describes how regimes permit and guide limited online criticism to act as a pressure-release valve. In China, for example, state actors seed stories critical of local officials while suppressing criticism of central leadership (Gainous et al., 2023). This approach redirects public frustration downward, projecting accountability at the margins while insulating elites. In this way, DDD suggests that regimes may permit dissent where it is least threatening—often redirecting criticism toward local actors while insulating central elites. Rather than suppressing all criticism, modern autocrats may channel it into “safe” spaces, creating a hierarchical pattern in which dissent is visible but carefully bounded. This expectation of a structured local–central gradient provides a clear empirical implication that we test directly below.

These developments raise important questions about how citizens encounter government-critical content in hybrid media systems—especially where online and offline channels differ dramatically in the range and tone of political information. Do citizens’ curated networks amplify or mute exposure to dissent? How do digital behaviors relate to engagement with legacy media, particularly television, which remains the dominant broadcast platform in many hybrid regimes? And how do individual political dispositions, such as interest, and macro-level structures, such as media freedom, moderate these relationships?

To investigate these questions, we analyze original survey data from three post-Soviet countries with varying levels of media freedom: Georgia, Kyrgyzstan, and Kazakhstan. We focus on three key outcomes: (1) exposure to critical content on social media, (2) political news consumption on television, and (3) perceived exposure to criticism of government on TV. We examine how these outcomes are associated with network curation, digital dissent exposure, and political interest, while accounting for cross-national differences using country fixed effects and, in alternative specifications, variation in media openness. Our framework highlights how curated digital networks—environments intentionally constructed by users to prioritize like-minded content—may either reinforce or disrupt authoritarian containment strategies like DDD.

Our results reveal several important patterns. First, individuals embedded in curated digital networks report greater exposure to critical content on social media, higher political news consumption on television, and greater perceived televised criticism of the government. This pattern suggests that rather than passively insulating citizens, curated networks may function as conduits for dissent, directing attention across platforms and priming viewers to recognize even subtle criticism in state-controlled broadcasts. Second, political interest moderates some of these relationships: high-interest individuals embedded in curated networks are more likely to engage with and interpret media critically. Third, exposure to critical content on social media is associated with stronger relationships between curation and legacy media behaviors and perceptions.

Importantly, when we directly test the hierarchical implication of DDD—that criticism is strategically channeled toward local actors while shielding central authorities—we find limited evidence of a pronounced local–central gradient. Instead, digital dissent exposure is associated with perceptions of both local and central criticism. This pattern suggests that once critical information circulates through curated digital networks, it may diffuse upward as well as downward in citizens’ interpretations, weakening the hierarchical containment logic implied by DDD. Notably, we do not find strong country-level effects; rather, individual-level dynamics—including network curation, political interest, and digital exposure—play a more central role than national media system variation in shaping how citizens perceive and interact with political information. We elaborate on the implications of this finding for DDD in the results and conclusion sections.

Taken together, our findings suggest that the effectiveness of DDD strategies may depend not only on top-down manipulation of information flows but also on bottom-up processes of digital curation and individual engagement. Curated digital networks do not merely filter content; they are associated with greater cross-platform political awareness, especially when dissent circulates through them and users are motivated to follow those threads into legacy media contexts. In short, even limited cracks in the information wall may be widened by citizens’ own media habits. Understanding these dynamics is crucial for evaluating the evolving limits of authoritarian information control in the digital age.¹

2 Theoretical framework: curated digital networks, media freedom, and critical information exposure

2.1 Media freedom, information control, and directed digital dissidence

Media system openness fundamentally structures the circulation of dissenting political information. In free media environments, independent outlets act as watchdogs, airing criticism of those in power and facilitating pluralistic debate. By contrast, in authoritarian and semi-authoritarian systems, legacy media (especially television) typically serve as state propaganda organs or regime mouthpieces, severely limiting critical coverage. For example, Freedom House reports that in Kazakhstan virtually all major media are either state-owned or controlled by government-friendly business elites, leaving little room for independent journalism (Freedom House, 2019). Authoritarian regimes thus rely on tight control of legacy platforms like TV to push pro-government narratives and suppress dissent.

Importantly, modern autocracies have not confined their information management to outright censorship; they also employ strategies of directed digital dissidence (DDD) to channel and contain criticism. According to DDD theory (Gainous et al., 2023), autocratic governments may tolerate or even tacitly encourage limited online criticism as a “relief valve” for public frustrations, so long as such criticism is directed at safe targets and does not threaten the core leadership. For instance, the Chinese Communist Party allows internet users to criticize local officials or policies while swiftly suppressing challenges to top Party authorities. This calibrated tolerance stabilizes the regime by deflecting discontent away from central elites. Indeed, King et al. (2013) find that Chinese censorship is not aimed at silencing all negative speech; instead it targets posts that could spur collective action, while permitting substantial criticism that lacks mobilizing potential. In short, authoritarian information control now mixes repression with strategic redirection: limited dissent is orchestrated in digital spaces to vent grievances without undermining regime legitimacy.

This means that in semi-authoritarian contexts, flows of government-critical content are not entirely absent but are theorized to be carefully stage-managed, often with an implicit hierarchical boundary between permissible local criticism and restricted central critique. The baseline availability of dissenting information, especially on dominant platforms like television, remains low, forcing citizens who seek critical perspectives to turn to alternative channels. Whether they actually encounter such dissent, however, depends not only on these top-down controls but also on how individuals navigate and curate their own information streams.

2.2 Curated digital networks: echo chambers, filter bubbles, and selective exposure

Even when critical information is available in the system, individuals often curate their media exposure in ways that reinforce prior beliefs and filter out unwelcome views. We introduce the concept of curated digital networks to describe the personalized information environments individuals create through selective choice of sources, social connections, and algorithmic tailoring. Rather than treating echo chambers, filter bubbles, and selective exposure/avoidance as separate phenomena, we argue that these mechanisms are now deliberately and strategically combined by users to shape their digital experience. A critical distinction in our framework is between curation as a process and the informational environments that result from that process. Traditional accounts of echo chambers and selective exposure largely treat homogenous information environments as emergent outcomes of preference-consistent behavior or algorithmic filtering (Garrett, 2009; Bakshy et al., 2015). In contrast, we conceptualize curated digital networks as the product of intentional and iterative user behavior, in which individuals actively shape their information streams by learning how algorithmic systems respond to engagement signals.

As users have become more media literate and algorithm-aware, they increasingly engage in behaviors—liking, sharing, commenting, following—not only to signal preferences but to actively train algorithms to show them more of what they want. This is no longer a passive consequence of platform architecture but a purposeful act of self-curation. Traditional terms like “echo chamber” or “filter bubble” fail to capture this degree of user agency. By unifying these concepts under the term curated digital network, we emphasize how user intent and algorithmic design jointly structure political exposure. In our study, we capture both condition and intent within a single measure, combining indicators of perceived network homogeneity, selective avoidance, and algorithmically responsive engagement to reflect both the structure of information exposure and the intentional behaviors through which users shape that exposure.

Curated networks can either reinforce state narratives or subvert them. If a citizen’s network is homogeneously pro-regime, they are likely to filter out critical news (selective avoidance), inadvertently reinforcing state messaging (Garrett et al., 2013). Conversely, if one curates a network of opposition voices or diverse perspectives that reflect or reinforce critical predispositions (selective exposure), even a tightly controlled media system can be disrupted as pockets of permitted dissent are amplified (Garrett, 2009). Thus, curated networks may insulate some users, but they may also sensitize others to political critique—especially when dissent is present but subtle.

We expect that the structure of one’s curated digital network will influence not only exposure to dissent online (H1), but also how users consume and interpret political content across platforms (H2 and H3, developed below). Highly curated networks may motivate individuals to monitor state TV for alignment or contradiction, and may prime them to notice instances of televised criticism others overlook. In this way, curated networks may condition both behavior and perception, particularly under semi-authoritarian constraints. We therefore state our first hypothesis formally:

H1: Individuals embedded in more highly curated digital networks report greater exposure to government-critical content (dissent) on social media.

At the same time, we acknowledge a growing literature that questions the strength of echo chamber effects, particularly in high-choice media environments, suggesting that exposure to diverse viewpoints may be more common than previously assumed. Our framework builds on this insight by emphasizing that contemporary digital environments are not purely passive or algorithmically determined; rather, users actively shape their information streams, potentially amplifying both agreement and dissent depending on how those networks are curated.

Recent work on algorithmic awareness and user–algorithm interaction suggests that users are not purely passive recipients of personalized content and may develop practical understandings of how engagement shapes what they see (Eckles et al., 2016; Eyal, 2014; Karizat et al., 2021). For example, users routinely describe distinct, self-constructed feeds (e.g., “Politics TikTok” or “Comedy TikTok”) and engage in behaviors such as liking, commenting, or rapidly scrolling past content in order to refine what the algorithm delivers. In this sense, curation is not simply the outcome of exposure dynamics but an ongoing process of algorithmic co-production, in which users actively engineer their informational environments.

This distinction is central to our contribution. While echo chambers and filter bubbles describe the structure of information exposure, curated digital networks emphasize the agency and intentionality underlying that structure. This shift has important implications for political learning: curated environments may not merely insulate users from dissent, but may also amplify and direct attention toward specific forms of political information depending on how users shape their networks. Importantly, this perspective implies that the same mechanisms that can produce homogeneous information environments may also facilitate the amplification and recognition of dissent, depending on how users intentionally shape their networks.

2.3 Legacy vs. digital media exposure under authoritarianism

Digital and legacy media function differently in constrained information environments. Television remains the dominant legacy platform in many semi-authoritarian regimes, delivering tightly managed, pro-regime content. Under freer conditions, mainstream TV may incidentally expose audiences to opposing views (Chadwick, 2013), but in authoritarian contexts, television tends to exclude dissent. Citizens relying exclusively on state TV thus receive limited or no exposure to criticism of those in power.

Digital media, in contrast, present both opportunities and challenges. While they can expose users to critical content and oppositional voices (Gainous et al., 2015, 2019; Bekmagambetov et al., 2018), they also allow for the construction of ideologically narrow information environments. Thus, digital exposure does not guarantee pluralism. Some users engage widely, encountering counter-attitudinal views; others retreat into partisan silos (Pearce, 2014; Dubois and Blank, 2018). We argue that the impact of digital media on political learning depends on how individuals curate their exposure—and whether they use digital media to supplement or replace legacy news.

Our framework anticipates that digital dissent exposure will be associated with greater political news consumption on television (H2) and greater sensitivity to critical content (H3), particularly among those in curated networks. Individuals do not consume platforms in isolation. Rather, they build cross-media habits shaped by intent, predisposition, and opportunity. Stated formally:

H2: Individuals embedded in more highly curated digital networks report greater political news consumption on television.

H3: Individuals embedded in more highly curated digital networks report greater perceived exposure to government criticism on television.

2.4 Political interest as a moderator of information curation

Political interest shapes what people seek and how they respond to information. Those with high political interest tend to consume more news and are more likely to actively curate their information environments. However, interest can both broaden and narrow exposure depending on network structure (Stroud, 2010; Bruns, 2019; Iyengar and Hahn, 2009). Highly interested individuals embedded in curated networks may amplify their exposure to like-minded content or critical perspectives depending on their orientation.

In our framework, we expect political interest to moderate the effects of curation on all three outcomes: exposure to digital dissent (H4 on H1), cross-platform political news consumption (H4 on H2), and interpretation of legacy media (H4 on H3). Those high in political interest are more likely to pursue news, recognize critique, and engage across platforms. Among the disinterested, curation may not yield the same behavioral or perceptual consequences. We state this moderating expectation formally:

H4: Political interest moderates the associations described in H1–H3, such that the positive associations between curated digital networks and (a) exposure to digital dissent, (b) television news consumption, and (c) perceived televised criticism are stronger among individuals with greater political interest.

2.5 Incidental exposure in curated digital networks

Much political exposure online is incidental: users encounter dissent not through active searching but via algorithmically surfaced posts or peer sharing (Bennett, 2008; Fletcher and Nielsen, 2018; Tewksbury et al., 2001). This matters in constrained contexts where traditional media avoid criticism. Even those who report low exposure to dissent may be influenced by incidental encounters within their curated networks.

We expect that critical exposure on social media will amplify the behavioral and perceptual effects of curation (H5 and H6). That is, the influence of curated networks on television news consumption and perceived criticism should be stronger among those who are already exposed to digital dissent. In contrast, among those with limited critical exposure, curation may not trigger the same interpretive engagement. This final component of our model helps distinguish between curation as structure and curation as a pathway—reinforcing that what citizens do with their networks depends not only on their design, but on the flows that move through them. We state these final two hypotheses formally:

H5: Exposure to critical content on social media moderates the association between curated digital networks and television political news consumption, such that the association is stronger among individuals with greater exposure to digital dissent.

H6: Exposure to critical content on social media moderates the association between curated digital networks and perceived exposure to criticism on television, such that the association is stronger among individuals with greater exposure to digital dissent.

3 Data and methods

The survey we use to test our expectations targeted adults aged 18 and older across Georgia, Kazakhstan, and Kyrgyzstan, excluding individuals living in institutional environments or hard-to-access regions. Each national sample comprised a minimum of 1,200 participants: Georgia (n = 1233), Kazakhstan (n = 1207), and Kyrgyzstan (n = 1201). Data collection was conducted between late November and mid-December 2023, concluding by December 15th. Interviews were administered via Computer-Assisted Telephone Interviewing (CATI) systems, with samples generated through Random-Digit Dialing (RDD) methods that randomly produced phone numbers without stratification. Comprehensive quality assurance procedures, including back-checks and statistical verification, were employed to maintain data integrity. According to Reporters Without Borders (2024), Georgia exhibits substantially higher press freedom than Kazakhstan, with Kyrgyzstan occupying an intermediate position.

3.1 Imputation strategy

Missing data on survey indicators were addressed through multiple imputation. We generated 30 posterior-drawn replicate datasets under a missing at random (MAR) assumption using an iterative imputation procedure with Bayesian regression estimators. This approach follows standard recommendations for multiply imputed datasets (Little and Rubin, 2014). We use 30 imputations—rather than the conventional 5 to 10—to ensure stable estimation and to minimize Monte Carlo error in the imputation process. All imputations were generated using fixed random seeds to ensure full reproducibility of the imputation process. To facilitate consistency across analyses, we construct a single dataset by averaging across the 30 imputed datasets after imputation and index construction. This approach allows all descriptive statistics, figures, and regression models to be based on the same underlying data, ensuring direct comparability across empirical results. While it is possible to achieve similar consistency by fixing the imputation seed and reusing draws, averaging provides a straightforward and transparent solution, particularly when integrating descriptive and multivariate analyses [see Gainous and Wagner (2023) for discussion]. All models reported in the paper are estimated on this averaged imputed dataset with heteroskedasticity-robust standard errors.²

A heatmap of missingness patterns is presented in the Supplementary Material, which shows that item nonresponse was generally low across key variables, though modest imputation was required for income and selective avoidance measures. Because the imputation model assumes approximate normality, some imputed values for ordinal variables (e.g., political interest) fall between scale categories; these values are retained to preserve information and avoid discretization bias.

3.2 Variables

We constructed composite indices (scaled 0–1) and utilized single indicators to capture the main independent variables (see the Supplementary Material for the question wording on each item):

• Curated digital network index: This additive scale combines respondents’ self-reported political like-mindedness of social media networks, frequency of disagreement on social media (reverse-coded), and frequency of selective avoidance behaviors (blocking, unfriending, leaving groups, unsubscribing)—all on ordinal Likert-type scales. All items were rescaled from 0–1 (preserving original intervals), summed, and then rescaled again to a continuous 0–1 scale. We follow the same process for all variables to ensure they are on comparable scales, facilitating clearer interpretation and presentation. Higher values indicate more curation of like-minded content. Cronbach’s alpha for the index is 0.77.³
• News source balance: Respondents were asked which sources they primarily get news from, choosing among five options ranging from “only traditional sources” (e.g., television/newspapers/radio) to “only digital sources” (e.g., internet/social media). These were recoded to range from 0 (only traditional) to 1 (only digital), with midpoint values reflecting an equal balance.
• Political interest: Respondents rated their interest in politics on a 4-point scale (very interested, somewhat interested, somewhat uninterested, or very uninterested). Responses were reverse-coded and rescaled from 0–1, where higher values indicate greater political interest.
• Media freedom: Each respondent was assigned a media freedom score based on their country’s standing in the 2024 Reporters Without Borders Press Freedom Index. These were normalized to range from 0.50 (least free) to 1 (most free), with intermediate values (e.g., 0.75) for middling cases. We began the scale at 0.50 to enhance interpretability in graphical displays.

The three dependent variables are:

• TV political news consumption: A measure capturing how often respondents consume political news via television, rescaled to a continuous 0–1 range.
• Exposure to critical news on TV: An index combining perceptions of seeing content critical of local and central government on television, also rescaled to range continuously from 0–1. The Cronbach’s alpha was 0.87.
• Exposure to critical news on social media: An additive index combining perceptions of seeing content critical of local and central government on social media platforms, rescaled to range from 0–1. This variable also appears as an independent variable in later models. The Cronbach’s alpha was 0.86.

3.3 Controls

We control for demographic variables known to influence media consumption and exposure: gender (0 = Female, 1 = Male), age (scaled 0–1), education level (scaled 0–1), household income (scaled 0–1), and urban residence (0 = Rural, 1 = Urban). These factors allow us to isolate the associations between curated networks, media use, and political interest and the dependent variables.

3.4 Analytical strategy

Given the small number of higher-level units (three countries), estimates of between-country variance are inherently imprecise. We therefore treat multilevel results as descriptive and rely primarily on the consistency of pooled, fixed-effects, and multilevel specifications. We initially estimated random-intercept models with random slopes for Curated Digital Network and News Source Balance; however, estimated country-level variance was minimal (see the Supplementary Material for full results).⁴ In the main models, we include country fixed effects by adding dummy variables for Kazakhstan and Kyrgyzstan (with Georgia as the reference category). This approach ensures that all reported associations reflect within-country relationships and are not driven by cross-national differences. As an additional specification, we also estimate models including media freedom as a country-level covariate instead of country fixed effects.⁵

We estimate three sets of ordinary least squares (OLS) regression models. OLS is a standard technique for estimating the linear relationship between a continuous dependent variable and a set of predictors; it identifies the coefficients that minimize the sum of the squared differences between observed and model-predicted values of the outcome. Because all of our dependent variables are rescaled to a continuous 0–1 range, each estimated coefficient can be read as the expected change in the outcome (on that 0–1 scale) associated with a one-unit change in the predictor, holding the other variables in the model constant. First, Table 1 presents six models predicting exposure to critical content on social media, television political news consumption, and perceived exposure to criticism on television. These models include additive specifications and interactions between Curated Digital Network and Political Interest to test H1 through H4. Second, Table 2 presents two models that introduce Exposure to Critical News on Social Media and its interaction with Curated Digital Network to test H5 and H6. Third, Table 3 disaggregates televised criticism into local and central targets and models the local–central difference to evaluate the hierarchical implication of DDD. All models include country fixed effects, with Georgia as the reference category, ensuring that the estimated relationships reflect within-country variation. We begin below with descriptive statistics for all key variables to contextualize the results that follow.⁶

Table 1. Models predicting exposure to critical content on social media (Models 1–2), TV political news consumption (Models 3–4), and perceived critical content on TV (Models 5–6).

Each pair of columns predicts one dependent variable; within each pair, the left column is the additive baseline model and the right column adds the Curated Network × Political Interest interaction (“Echo × Interest”). Coefficients are estimated on the averaged imputed dataset; HC1 robust standard errors in parentheses. Models include country fixed effects (Kazakhstan and Kyrgyzstan, with Georgia as the reference category). ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01.

Table 2. Models predicting TV news consumption and perceived critical content (SM critical × curated network interaction).

Coefficients are estimated on the averaged imputed dataset; HC1 robust standard errors in parentheses. Models include country fixed effects (Kazakhstan and Kyrgyzstan, with Georgia as the reference category). ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01.

Table 3. Testing directed digital dissidence: local vs. central criticism on television.

Coefficients are estimated on the averaged imputed dataset; HC1 robust standard errors in parentheses. Models include country fixed effects (Kazakhstan and Kyrgyzstan, with Georgia as the reference category). ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01.

4 Results

4.1 Descriptive findings

There is clear evidence of meaningful patterns across the key predictors, as shown in Figure 1. The distribution of the Curated Digital Networks index is heavily skewed toward lower values, with a clustering near zero on the 0–1 scale. This indicates that while many respondents report relatively open networks with frequent exposure to disagreeable views, a substantial portion nonetheless curate more like-minded networks. The distribution suggests significant individual-level variation in how much social and informational homogeneity citizens experience.

The News Source Balance index, capturing the extent to which respondents rely on digital versus legacy media for news, is skewed toward heavy reliance on digital sources.⁷ This distribution is consistent with a shift in media diets and has implications for the analyses below, as digital sources provide comparatively less constrained alternatives to state-influenced legacy media. Here, however, we still observe a non-trivial share of respondents who report reliance on traditional sources or a more balanced news diet. Overall, this distribution underscores that the divide between traditional and new media ecosystems leans strongly toward digital reliance, especially salient in post-Soviet contexts where legacy media remains more tightly constrained while digital alternatives are comparatively less controlled.

Political Interest, shown in the lower-left panel of Figure 1, exhibits a relatively spread distribution, with peaks at both low and moderately high ends and a large concentration around the midpoint. This suggests a segmented population in terms of political attentiveness: a clear presence of politically disengaged respondents (very or somewhat uninterested) alongside a substantial share who report at least some interest.

Finally, the bottom panel of Figure 1 reveals clear cross-national differences in Media Freedom. Kazakhstan is the least free, Kyrgyzstan occupies an intermediate position, and Georgia has the highest media freedom score. Figure 2 presents the country means for the dependent variables. Notably, TV Political News Consumption (top-left panel) does not differ substantially across countries, with overlapping confidence intervals across Kyrgyzstan and Kazakhstan and only modestly higher levels in Georgia. Levels are generally low across all three contexts, consistent with the descriptive evidence that most respondents rely heavily on digital sources, suggesting that television is no longer the dominant platform for political news in the region.

By contrast, Perceived Exposure to Critical News on TV (top-right panel) shows statistically significant differences across all countries: Georgia scores highest, followed by Kyrgyzstan and Kazakhstan. Exposure to Critical Content on Social Media (bottom-left panel) shows some variation, with Georgia again leading, but the relative means for Kyrgyzstan and Kazakhstan are not statistically distinguishable. Taken together, these descriptive findings suggest that while television consumption is relatively stable across contexts, perceptions of government criticism vary more sharply—foreshadowing our core argument that perceptions of criticism are closely linked to digital environments and political predispositions.

4.2 Regression results and hypothesis tests

Table 1 presents the main regression results. The six models are organized as three pairs, one pair for each dependent variable: Models 1–2 predict exposure to critical content on social media, Models 3–4 predict television political news consumption, and Models 5–6 predict perceived exposure to criticism on television. Within each pair, the first (odd-numbered) model is an additive baseline specification that includes only the main effects of the predictors, while the second (even-numbered) model adds the interaction between Curated Network and Political Interest (labeled “Echo × Interest” in the table) in order to test the moderating expectation in H4. The two members of each pair thus estimate the same outcome, differing only in whether the moderation term is included. We find strong support for H1. The Curated Digital Network Index is positively and significantly associated with exposure to critical content on social media (Table 1, first column).^8,9 This indicates that individuals embedded in ideologically homogeneous networks, whether through selective connection or algorithmic reinforcement, report greater exposure to government-critical perspectives online. Consistent with the inclusion of country fixed effects, respondents in Kazakhstan and Kyrgyzstan report systematically lower levels of exposure to critical content and political news consumption compared to Georgia (the reference category). These differences reflect cross-national variation but do not alter the substantive relationships between curated networks and the dependent variables. More importantly for our theoretical expectations, curated digital networks are associated with higher reported levels of dissent exposure in these environments.

Next, the positive and significant coefficient for Curated Digital Network in the model predicting TV Political News Consumption is consistent with H2. Individuals embedded in more curated networks report higher levels of political news consumption on television, even though television remains largely state-controlled. This pattern aligns with our argument that digital dissent exposure may correspond with cross-platform engagement rather than substitution, with individuals monitoring legacy media for confirmation, contradiction, or additional political information.

We also find clear support for H3. Individuals in curated digital networks with higher exposure to critical information on social media are significantly more likely to report perceiving government criticism on television. This suggests that curated networks are associated not only with differences in media consumption but also with differences in how individuals interpret legacy content. Exposure to dissent on social media appears linked to greater recognition of subtle forms of televised criticism, even when such content is limited. In this way, curated digital networks and critical online flows are associated with variation in the perceptual filter citizens apply to state media.

Turning to H4, the results are more nuanced. Political interest moderates some of these relationships, though the effects are not consistently statistically significant across outcomes. However, contrary to our expectations, the interaction between political interest and curated networks does not significantly influence exposure to critical content on social media. This partial support for H4 suggests that while political interest is associated with cross-platform engagement and interpretation, it does not appear to structure digital dissent exposure itself to the same degree. This may indicate that network structure reflects broader social or algorithmic dynamics that operate somewhat independently of individual political attentiveness.

The interactions displayed in Figure 3 further clarify this dynamic. Each panel plots predicted outcomes on the 0–1 scale (y-axis) across the full range of the Curated Digital Network index (x-axis; 0 = least curated, 1 = most curated), with separate lines for respondents at lower (blue) versus higher (orange) levels of political interest and shaded bands showing 95% confidence intervals; the left panel presents predicted TV political news consumption and the right panel predicted perceived exposure to government criticism on TV. As the figure shows, political interest strengthens the relationship between curated networks and both television news consumption and sensitivity to televised criticism, but it does not substantially alter digital dissent exposure. Thus, interest appears more consequential at the stages of interpretation and cross-platform engagement than at the initial stage of digital network formation. The predicted patterns are nevertheless substantively interpretable. Moving from the lowest to highest level of curated networks corresponds to an increase of approximately 0.20 points in predicted television news consumption among low-interest respondents, compared to a smaller change among high-interest respondents. A similar pattern emerges for perceived criticism on television, where the slope of curation is markedly steeper among those with lower political interest, while high-interest respondents begin at higher predicted levels and change less steeply. These differences underscore that the interaction reflects meaningful variation in predicted outcomes across the observed range of the curation index, rather than merely statistical significance, and illustrate the partial support for H4: political interest conditions how citizens interpret and consume television content more than it alters the overall positive relationship between curation and the television outcomes.

H5 and H6 are tested using models that include an interaction between curated digital networks and exposure to critical information on social media (Table 2, Figure 4). These interactions are negative and statistically significant. In Figure 4, each panel plots predicted outcomes on the 0–1 scale (y-axis) across the full range of the Curated Digital Network index (x-axis), with separate lines for respondents with lower (blue) versus higher (orange) exposure to critical content on social media and shaded bands showing 95% confidence intervals; the left panel presents predicted TV political news consumption and the right panel predicted perceived exposure to government criticism on TV. Substantively, respondents with high exposure to critical content on social media report higher levels of political television news consumption and perceived government criticism across levels of network curation. The negative interaction indicates that the marginal effect of curated networks is strongest among those with lower levels of exposure to critical content on social media, while those already highly exposed exhibit elevated outcome levels regardless of network intensity; however, this does not reverse the overall pattern. Rather, exposure to digital dissent is independently associated with higher levels of television news consumption and perceived televised criticism. Among respondents with low exposure to critical content on social media, moving from low to high levels of network curation corresponds to an increase of approximately 0.25 points in predicted television news consumption. In contrast, among those already highly exposed to digital dissent, predicted consumption levels remain elevated across the curation spectrum, indicating that online critical exposure itself is associated with higher engagement regardless of network intensity. A parallel pattern appears for perceived criticism on television, where exposure to digital dissent is associated with consistently higher predicted values across levels of curation.

Importantly, when we disaggregate criticism into local and central targets to directly test the hierarchical implication of Directed Digital Dissidence (DDD), we find no evidence of a systematically structured local–central gradient. Although local criticism is slightly more prevalent on average, the difference is small and not structured by curated networks. Moreover, exposure to critical content on social media is positively associated with perceptions of both local and central criticism on television. These findings suggest that once dissent circulates through curated digital networks, it is associated with broader perceptions of criticism across hierarchical targets rather than confinement to “safe” local actors. In this respect, the evidence qualifies the containment logic implied by DDD.

To directly evaluate the hierarchical implication of Directed Digital Dissidence (DDD), Table 3 disaggregates perceived criticism on television into local and central targets and models the local–central difference. If DDD operates through structured containment, criticism should be disproportionately directed toward local actors, with limited spillover to central authorities. The results provide only limited support for such a hierarchical pattern. Curated digital networks are positively and significantly associated with both local and central criticism on television, and the magnitude of these associations is similar across the two columns. Importantly, the coefficient for curated networks in the gradient model (local minus central) is small and statistically insignificant, indicating no evidence that curation is differentially associated with local versus central criticism on television.

Exposure to critical content on social media is also positively associated with perceived criticism of both local and central authorities. Notably, the association is slightly stronger for central criticism, and the gradient coefficient is negative and statistically significant. This pattern indicates that higher levels of digital dissent exposure are associated with a narrowing—and in some cases reversal—of the local–central distinction in perceived televised criticism. Rather than aligning with a strictly “safe” channeling of criticism toward local actors, digital exposure is associated with broader perceptions of criticism across hierarchical levels. In alternative specifications using media freedom instead of country fixed effects, we observe a similar pattern: it is positively associated with both local and central criticism, but more strongly with central criticism, resulting in a negative and significant gradient coefficient. Thus, greater openness in the media environment is associated with increased perceived criticism overall, not only criticism concentrated at the local level.

Taken together, these findings suggest that while curated digital networks and digital dissent exposure are associated with higher perceived criticism on television, they do not align with a strongly structured local–central containment pattern. Instead, once dissent circulates through curated digital networks, perceived criticism appears to extend across hierarchical targets rather than remaining confined to “safe” local actors. This result complicates the containment expectations implied by DDD and points toward a more porous and less tightly bounded information environment than the theory might predict. We turn next to the broader implications of this finding for authoritarian information control.

As a further robustness check, we estimated alternative specifications using media freedom in place of country fixed effects (Supplementary Table 2). The substantive findings remain unchanged. Curated digital networks remain positively associated with exposure to critical content on social media, political news consumption on television, and perceived televised criticism. Likewise, exposure to critical content on social media and political interest retain their general pattern of association. These results indicate that the findings are robust to alternative ways of accounting for cross-national variation.

Before turning to the broader implications of these findings, it is important to clarify the limits of causal inference in this study. All analyses are based on cross-sectional survey data and therefore identify statistical associations rather than directional causal effects. Although the models control for political interest, media use patterns, and demographic characteristics, it remains possible that individuals predisposed toward dissent both curate more homogeneous networks and report higher exposure to critical content. Likewise, politically engaged respondents may simultaneously consume more media and report greater perceived criticism. As such, the results should be interpreted as evidence of robust associations consistent with our theoretical expectations, rather than definitive evidence of causal pathways. Longitudinal or experimental designs would be required to adjudicate directionality more conclusively. Moreover, reverse causality cannot be ruled out; individuals who are already more attentive to political dissent may be more likely to curate like-minded networks and to notice criticism across platforms.

5 Discussion

Our findings provide consistent evidence that the intersection of personalized media environments and structural media conditions is associated with political information exposure in semi-authoritarian contexts. Specifically, curated digital networks—even when intentionally constructed through selective following and engagement—do not merely coincide with insulated echo chambers; rather, they are associated with greater reported exposure to government-critical perspectives, particularly when dissent circulates online. While this exposure may be filtered and partial, it appears to coincide with broader political engagement and differences in how citizens interpret state media content.

Crucially, our results indicate that political interest amplifies some—but not all—of these dynamics. Those who are politically interested are more likely to act on the information they encounter across platforms. Interest strengthens the relationship between curated digital networks and both television news consumption and the perception of televised criticism. However, interest does not significantly moderate whether one is embedded in a curated digital network that exposes them to dissent in the first place. This pattern is consistent with the possibility that intentional digital behaviors such as selective unfollowing or blocking are habitual or socially structured rather than purely ideologically strategic.

Importantly, we also show that exposure to critical content on social media is associated with both political behavior and perception in legacy media spaces. Our interactive models indicate that citizens embedded in curated digital networks and exposed to online dissent are especially likely to consume television news and perceive critical content within it. In contrast, those in curated networks with limited exposure to dissent on social media are less attuned to criticism on television. This supports the idea that digital experiences shape not only what people see, but how they process state-controlled narratives.

At the same time, our direct test of Directed Digital Dissidence (DDD) yields a qualified conclusion. While DDD theory anticipates a structured pattern in which criticism is channeled primarily toward local actors while shielding central authorities, we find only a modest local–central difference in reported exposure. Moreover, curated digital networks are not associated with a stronger hierarchical gradient. Instead, digital dissent exposure is associated with perceptions of both local and central criticism. These findings suggest that once critical information circulates through curated networks, it may diffuse across hierarchical targets rather than remain confined to “safe” local criticism. In this respect, our results do not support a strong containment gradient and instead suggest that bottom-up digital dynamics can complicate hierarchical information control more than DDD alone would predict.

This study is not without limitations. First, because our data are cross-sectional, we cannot determine whether exposure to dissent causes citizens to change their media behavior or whether those predisposed to dissent seek out critical content. We address this by controlling for political interest and media use patterns, but experimental or longitudinal data would offer stronger causal leverage. Second, our key variables rely on self-reported behavior and perceptions, which are inherently subject to recall bias and social desirability effects. Lastly, although we account for cross-national differences using country fixed effects and alternative specifications with media freedom, the small number of countries limits strong cross-national inference; our results primarily reflect individual-level dynamics rather than structural country differences.

Nonetheless, these findings offer valuable insights for both scholars and practitioners. For researchers, our results reinforce the need to conceptualize echo chambers not merely as information silos but as dynamic, cross-platform political environments. Future work might explore how the interplay between interest, network curation, and dissent exposure unfolds in other hybrid or democratic contexts. For practitioners working in constrained media systems, our findings suggest that digital platforms may facilitate political awareness, but their effects likely depend on how users curate and engage with them.

Ultimately, fostering an informed citizenry in the digital age requires attention to both user-level behavior and institutional media structures. Curated digital networks do not simply insulate citizens from dissent; under certain conditions, they are associated with heightened attentiveness to political critique across platforms. Most importantly, the absence of a pronounced local–central gradient in our DDD tests suggests that hierarchical containment strategies may be more porous than previously assumed, at least in these cases. The core theoretical implication is that bottom-up curation can blur the vertical boundaries that DDD emphasizes: when dissent circulates widely in digital networks, it may not remain confined to local targets, and citizens may carry that critical orientation into their interpretation of legacy media. Understanding how digital curation interacts with authoritarian information management remains central to assessing the evolving limits of regime control.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. We used generative AI to assist with theorizing, Python code, writing and editing. We used it only as an assistive tool and take full responsibility for all content in the 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/fcomm.2026.1780602/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Eugenio Bagnini, University of Bologna, Italy

Reviewed by: Valia Kaimaki, Ionian University, Greece

Eglantina Kalluci, University of Tirana, Albania

Footnotes

¹The authors used generative artificial intelligence (AI) tools during the research and writing process. AI-assisted interactions were used to help refine theoretical framing through iterative discussion, generate and troubleshoot Python code for data analysis, and improve clarity and organization of the manuscript’s prose. All substantive decisions regarding research design, variable construction, modeling strategy, interpretation of results, and theoretical claims were made exclusively by the authors. The authors take full responsibility for the content of this article and confirm that AI tools were used as collaborative aids rather than as independent analytic agents.

²As a robustness check, we re-estimated all models using Rubin’s rules to pool estimates across the 30 imputed datasets. The results are substantively identical. Coefficient directions are consistent in over 99% of cases, and statistical significance classifications agree in over 96% of cases. Differences in coefficient magnitudes are minimal, with a mean absolute difference of 0.013. These findings indicate that the substantive conclusions are not sensitive to the choice of imputation pooling strategy.

³Cronbach’s alpha is a measure of internal-consistency reliability that summarizes how strongly the items in a multi-item scale intercorrelate—that is, the extent to which they appear to measure the same underlying construct. It ranges from 0 to 1, with higher values indicating greater consistency among items. By common convention, values at or above 0.70 are considered acceptable, values above 0.80 good, and values above 0.90 excellent. The alphas reported for our indices (0.77–0.87) therefore indicate acceptable-to-good internal consistency.

⁴The mixed-effects models yield very small country-level variance components (ICC ≈ 0), indicating limited clustering. In such cases, multilevel estimates converge toward pooled OLS. This is further supported by fixed-effects specifications, which yield substantively identical results.

⁵Because media freedom is constant within countries, it is collinear with country fixed effects and cannot be included simultaneously. Models using media freedom instead of country dummies yield substantively identical results, indicating that the findings are robust to alternative ways of accounting for cross-national variation. Media freedom is therefore interpreted as a country-level characteristic rather than an individual-level predictor.

⁶All data, code, and replication materials will be made publicly available via Harvard Dataverse upon publication.

⁷The short bars between the central peaks are imputed values. Recall that this indicator is ordinal.

⁸We describe a coefficient as positive or negative according to its sign: a positive coefficient means that higher values of the predictor are associated with higher values of the outcome, and a negative coefficient the reverse. Because every variable is scaled 0–1, a coefficient’s magnitude is the predicted change in the outcome (also on a 0–1 scale) associated with moving the predictor across its full range. We describe a coefficient as statistically significant when the estimate is large relative to its standard error, such that the probability of observing an association this large by chance if the true association were zero (the p-value) falls below a conventional threshold. Following standard practice, the asterisks in the tables classify significance into three ranges: ∗p<0.10, ∗∗p<0.05, and ∗∗∗p<0.01. A “positive and significant coefficient” therefore denotes an estimate with a positive sign whose p-value falls below at least the 0.05 level (two asterisks) unless otherwise noted.

⁹To assess whether the bundled Curated Digital Network index obscures substantively distinct effects of perceived homogeneity and selective avoidance, we re-estimated the models replacing the combined index with its two constituent components (echo index and avoidance index). Results (available in the Supplementary Material) indicate that both components are independently and positively associated with exposure to critical content on social media and perceived exposure to criticism on television. For political news consumption on television (see below, Table 2), perceived network homogeneity remains statistically significant, while selective avoidance is weaker but in the expected direction. These patterns suggest that the bundled index captures a coherent underlying construct rather than masking divergent effects. Accordingly, we retain the combined Curated Digital Network measure in the main models for parsimony and conceptual clarity.