Introduction

In the digital public sphere, diversity of viewpoints and rationality of discussion are widely recognized as two core features of public deliberation, serving as important mechanisms for promoting healthy democratic discourse (Dryzek, 2000; Habermas, 1996). Diversity emphasizes the inclusion of different opinions and perspectives in the deliberative process, helping to break information echo chambers and reduce the emergence of extreme positions (Mutz, 2006); rationality advocates for providing reasons, evidence, and logical arguments to support one’s viewpoints, thereby facilitating information sharing and cognitive updating in discussions (Stromer-Galley, 2007).

However, today’s online space has witnessed the rapid rise of populism. In China in particular, although the meritocratic political system has to some extent constrained the emergence of populist politicians and effectively precluded top-down populist mobilization, a form of bottom-up populist expression continues to proliferate on the internet (Ma, 2015). Chinese online populism is characterized by grassroots political narratives, with ordinary netizens leveraging anonymity to launch collective criticism against elite misconduct and perceived threats from “the other” (He et al., 2021; Miao et al., 2020). Here, “the elite” refer to those who ostensibly speak on behalf of the people but fail to genuinely represent their interests, having lost the sense of “paternalistic responsibility” (Miao et al., 2020). “the other” are those perceived as threatening societal or collective interests, such as Western countries or “white left” ideologies (Zhang, 2020; Zhang, 2022), reflecting Chinese netizens’ strong exclusionary attitudes and the defense of mainstream values. Thus, anti-elitism and nationalism together form the fundamental tone of Chinese online populism.

The extremely low threshold for participation on Chinese social media has led to the emergence and fermentation of numerous hotly debated topics that are permeated with the aforementioned populist tendencies. For instance, the “Driving a Mercedes into the Forbidden City”1 incident triggered intense public anger toward elite privilege and wealth (He et al., 2025b); similarly, discussions surrounding the “996” work schedule are filled with resistance to excessive overtime and calls for the protection of workers’ rights. There is also the case of the public outcry over foreign brands ceasing to use Xinjiang cotton in their products2 (Tao et al., 2025). However, current communication studies on such populist issues mostly focus on the discursive construction and logic of populist discourse within individual topics (He et al., 2025a; He et al., 2025b; Tao et al., 2025; Zhang & Schroeder, 2024), while there remains a lack of attention to how these populist discourses actually diffuse in the online sphere.

Whether diversity and rationality—two essential elements of deliberation—can curb the diffusion of populist discourse is the central question of this study. When diversity is present, the discussion space accommodates heterogeneous voices, thereby depriving populist discourse—which heavily relies on singular positions and adversarial constructions—of fertile ground for spreading (Sunstein, 2001; Cinelli et al., 2021). Likewise, when discussions are grounded in rationality, participants are more likely to engage with issues prudently and are less susceptible to emotional mobilization, thus hindering the proliferation of populist discourse (Rauchfleisch & Kaiser, 2021).

To examine this relationship, this study integrates computational analysis with traditional statistical testing. First, ten highly influential populist topics from Chinese social media were selected, and all related discussion threads from Zhihu—a major Chinese Q&A platform—were systematically collected as the research corpus. Next, a pre-trained large language model was employed to measure the two key predictor variables: diversity and rationality within the discussions. The number of comments and likes received by each thread were used as quantitative indicators of the extent of “diffusion.” Finally, regression analysis was conducted to explore the relationships among diversity, rationality, and the diffusion of populist discussions, thereby addressing the central research question.

This study makes two primary contributions: first, it deepens the understanding of the applicability and limitations of deliberative democratic theory in the context of non-Western digital platforms, expanding the conceptualization of diversity and rationality; second, it provides a theoretical basis for understanding the diffusion mechanisms of online populist discussions and offers insights for platform governance in China.

Theoretical Foundations of Public Sphere Deliberation

The public sphere, conceived by Habermas (1962/1989) as the central arena for democratic discourse, posits rational, equal, and non-coercive communication as the foundation for deliberation on public affairs. However, the ascent of digital media has presented profound challenges to this idealized model (Dahlberg, 2001). While online platforms have eliminated spatial constraints and enabled mass citizen participation, they have also ignited renewed debate over whether high-quality, rational deliberation is achievable in these new environments (Papacharissi, 2002). Empirical research by Brundidge (2010) further reveals that even on open social media platforms, users tend to engage with like-minded perspectives, reinforcing public sphere fragmentation.

In institutional studies of populism, scholars note that shifts in populist discourse in the United States and Europe since approximately 2015 are closely linked to structural transformations and institutional crises (Brubaker, 2017; Jansen, 2011). The evolving role of the media has been a particularly salient factor in these crises. The media landscape has not only altered professional journalism but has also fundamentally reshaped the modes of citizen engagement in political communication.

A review of the literature on populist diffusion identifies several key drivers since the 1990s: the impacts of globalization and supranational integration, the media’s role in amplifying the visibility of populist actors, and the strategic responses of mainstream political parties have all contributed significantly to the growing prevalence of populist discourse (Manucci & Weber, 2017).

Against this backdrop, our study focuses on two critical variables in the digital public sphere—diversity and rationality—to examine their influence on the diffusion of populist discourse. Grounded in deliberative theory, this research investigates the communicative features of online discussions, focusing specifically on how viewpoint diversity and rationality shape the propagation of populist discourse in the digital realm.

Diversity and Populism

Viewpoint diversity in the public sphere is considered a cornerstone of effective democratic deliberation (Dryzek, 2000). As a core deliberative variable, diversity not only embodies the normative democratic ideal of including heterogeneous voices but also shapes the quality of knowledge integration and social understanding that deliberation can produce (Dryzek, 2000).

In a comparative study of political forums, Janssen and Kies (2005) identified three structural variables that determine the quality of online deliberation: the forum’s communicative architecture, the political culture and ideological leanings of its participants, and the distinction between “strong” and “weak” publics. Their framework suggests that the deliberative function of the public sphere is only realized when platform design actively fosters encounters with heterogeneous viewpoints and creates a balanced participatory arena. This underscores that platform architecture itself profoundly shapes whether diversity is achieved and, consequently, whether it can serve as a counterbalance to the monolithic narratives often propagated by populism.

Deliberative diversity is not merely an aggregation of different opinions; it requires fostering cognitive engagement and social understanding at the intersection of similarity and difference. Analyzing discussion preferences, Morey, Kleinman, and Boukes (2018) found that dialogues featuring both a shared identity and divergent viewpoints are more effective at promoting learning, attitude change, and political tolerance. They argue that such conversations provide social support while mitigating excessive confrontation, creating an optimal environment for deliberation, compromise, and democratic norm-building. This “commonality within diversity,” they suggest, may be an effective means of countering the polarizing narratives characteristic of populist discourse.

Theoretically, exposure to a diversity of opinions is thought to counteract echo chambers and mitigate political polarization (Mutz, 2006). In practice, however, the algorithmic systems of digital platforms often create “echo chambers” (Sunstein, 2001) that reinforce in-group homogeneity by limiting exposure to divergent viewpoints. The concept of an echo chamber posits that users experience ideological isolation within their online media environments (Garrett, 2009). Indeed, empirical research confirms the prevalence of homophily and opinion segregation across online communities (Colleoni, Rozza & Arvidsson, 2014). Similarly, the “filter bubble” concept describes how personalized algorithms can immerse users in an increasingly narrow ideological space (Pariser, 2012, pp. 101–139). These filter bubbles, in turn, can foster the formation of more isolated and fragmented online communities (Borgesius et al., 2016).

These homogenized information environments provide fertile ground for the dissemination of populist discourse, which is often one-sided, simplistic, and emotionally charged (Mudde, 2004). This trend stands in contrast to the findings of Mutz (2006), who, drawing on data from the American National Election Studies (ANES), demonstrated that regular exposure to heterogeneous viewpoints can effectively ’temper extreme political attitudes. Nevertheless, algorithmic curation on digital platforms has intensified the effects of echo chambers and filter bubbles (Pariser, 2011), thereby further constraining the diversity of information available to users.

Specifically, a lack of diversity directly erodes the public sphere’s resilience to populist narratives. In the absence of viewpoint plurality, discussion spaces become susceptible to “group polarization” and “cognitive resonance,” fostering an environment ripe for the propagation of hostile populist rhetoric (Sunstein, 2001; Cinelli et al., 2021). For instance, Bobba and Hubé (2021) found that Italian and French political forums with low viewpoint diversity were easily exploited by populists who simplified complex issues into a “people vs. elites” binary to manipulate public emotions. Similarly, an analysis of 2016 US election data from Twitter showed that greater exposure to ideologically consonant content correlated with an increased propensity to share populist and false news (Guess et al., 2018). Furthermore, a cross-platform analysis of Facebook, Twitter, and YouTube revealed that right-wing populist discourse diffused more effectively within homogeneous communities, where closed information loops amplify antagonism and social hostility (Cinelli et al., 2021).

Diversity is thus considered a critical bulwark safeguarding public deliberation from extremist tendencies (Dryzek, 2000). Research consistently shows that when pluralistic voices are absent, populist discourse more easily reinforces social divisions through its binary “people versus the elite” narrative (Waisbord, 2018). Consequently, a lack of diversity not only undermines the rational foundations of the public sphere but also directly facilitates the diffusion of populism.

Rationality and Populism

Within public sphere theory, rationality is considered the cornerstone of deliberative democracy (Habermas, 1996). As Stromer-Galley (2007) argued, rational deliberation transcends mere opinion expression, requiring participants to provide evidence, cite external sources, and offer reasoned responses to counterarguments. A rational discussion is therefore characterized by logical argumentation, empirical grounding, engagement with dissent, and thematic coherence (Friess & Eilders, 2015).

In a systematic review, Friess and Eilders (2015) found that most assessments of online deliberative quality treat rationality as a core dimension, typically operationalized via three indicators: justification of claims, topical relevance, and use of external evidence. These criteria are not merely constitutive of deliberation but also influence key outcomes, including information gain, cognitive updating, and attitude modification.

In practice, however, online platforms often fall short of these rational standards. For example, Rowe’s (2015) comparative study of news comments, forums, and social media revealed that discussions on social media tend to be more affective, less evidence-based, and more prone to topic drift. Such an environment, characterized by low rationality, becomes susceptible to the spread of populist discourse.

A lack of rationality not only impairs individual judgment but also weakens a platform’s capacity for opinion integration. As Delli Carpini, Cook, and Jacobs (2004) highlighted in their meta-analysis, rational discussion serves a vital educative function, enabling citizens to gain a comprehensive understanding of issues and refine their views through collective discourse. When this function is compromised, populist discourse can more easily sway public opinion by mobilizing emotion rather than evidence, thereby bypassing the “deliberative checkpoints” of democratic processes.

The affective nature of many online discussions often subverts rational dialogue (Papacharissi, 2015). Populist discourse excels in this environment by leveraging emotions like fear, resentment, and identity-based appeals to offer simplified explanations for complex social problems (Norris & Inglehart, 2019). This approach de-emphasizes the normative value of facts, logic, and evidence in public discourse (Waisbord, 2020). Consequently, research indicates that when discussions lack factual grounding and critical reflection, populist narratives are more likely to prevail (Benkler et al., 2018).

Moreover, platforms’ algorithmic incentives often amplify affective interaction patterns, boosting the visibility and influence of non-rational discourse (Guess et al., 2018). For instance, a content analysis of Swiss political Facebook groups by Kaiser and Rauchfleisch (2019) found that emotional and moralized language significantly increased post engagement, while rational, evidence-based arguments were less likely to achieve widespread dissemination.

Cross-national research corroborates these trends. For instance, Törnberg (2018), employing network-based sentiment and social network analysis, examined far-right communities across Europe and found that their populist discourse functions to solidify in-group identity while marginalizing dissenting views. This is achieved through provocative and emotionally charged narrative frames, such as the invocation of a “cultural crisis” or the “loss of national identity.” This non-rational mode of dissemination is, in turn, amplified by algorithmic recommendations (Bakshy, Messing & Adamic, 2015).

Focusing on online political communication, Galpin and Trenz (2019) introduced “participatory populism” to describe how non-representative user groups actively delegitimize democratic institutions online. This participation style cultivates a collective voice defined by its staunch rejection of mainstream democratic politics. Analyzing user comments during the 2014 European Parliament elections in Germany and the UK, they demonstrated how online news engagement fostered this participatory populism, with users expressing intense negativity toward established institutions and political actors.

Furthermore, Rauchfleisch and Kaiser (2021) showed that during crises like the COVID-19 pandemic, populists exploited the climate of fear and uncertainty to intensify their emotional communication strategies. By mobilizing antagonistic sentiments, these tactics further marginalized rational policy discussions in the public sphere.

In summary, the levels of diversity and rationality within digital deliberation are critical in shaping the diffusion of populism. A lack of diversity can foster information silos and extremist views, while an erosion of rationality allows emotive, non-rational populist narratives to dominate. Building on these theoretical foundations, this study investigates how diversity and rationality manifest in online discussion spaces and how they jointly influence the spread of populist discourse.

Research Questions

Classical and contemporary populism research has been criticized for a Western-centric bias, often overlooking its manifestations in non-Western contexts (Tugal, 2021). This study addresses this gap by situating its analysis in the context of Chinese online populism.

Chinese social-media populism has become a pivotal meeting point for research on digital nationalism, political communication, and global populism. Early conceptual work treats this phenomenon as a concrete manifestation of digital nationalism, highlighting how Chinese netizens employ memes, hashtags, and cross-platform coordination to participate in public debates and to re-imagine both the nation and its “others” (Schneider, 2018). Schneider (2018) argues that such online practices translate macro-political narratives into personalized displays of identity and emotion, thereby intertwining nationalist and populist logics.

Large-scale data analysis provides a macro lens on these dynamics. Comparing more than 100,000 posts on Weibo and Twitter, Zhang and Schroeder (2024) show that “us-versus-them” narratives on Chinese-language platforms tend to target foreign actors, whereas overseas Chinese-language users focus more on domestic governance issues. Topic- and sentiment-model results from Chen et al. (2020) further indicate that COVID-19 conspiracy and rebuttal texts framed in explicitly nationalist terms attract markedly higher engagement. Together, these findings suggest that identity-laden and affective cues significantly boost message diffusion on Chinese social media.

Information control also shapes how digital populism unfolds. King, Pan and Roberts’ (2013) seminal study demonstrates that China’s censorship regime often permits moderate criticism of the government but rapidly removes content deemed likely to facilitate offline collective action. At the same time, official accounts occasionally adopt populist rhetoric in external communication to reinforce state positions and cultivate domestic consensus (Chen, 2023). The interaction of “bottom-up emotions” and “top-down guidance” thus forms a distinctive ecology for Chinese digital populist expression.

Despite mounting interest in Chinese social-media populism, scholarship has yet to clarify how diversity, rationality, and diffusion interact within a single platform ecosystem. Most studies rely on one-off events, leaving unexplored whether—and how—shifts in viewpoint heterogeneity and argumentative quality over time shape the visibility of populist claims. Moreover, existing work often conflates algorithmic ranking with community self-organization, making it difficult to separate the effects of platform design from those of user interaction in amplifying emotional content and muting reasoned debate. Finally, little empirical attention has been paid to the moderating roles of commercial influencers or counter-public voices in either accelerating or curbing populist spread. These gaps underscore the need for a platform-level, longitudinal approach that can capture sustained, threaded exchanges and disentangle technical from communal forces.

Our empirical site is Zhihu, a prominent knowledge-sharing platform in China where users generate threaded discussions by posting questions and answers. Many of these threads revolve around populist themes, attracting significant user engagement and achieving broad diffusion.

This study is therefore motivated by a central question: can the communicative dynamics within a platform like Zhihu act as a community-based mechanism to shape the diffusion of populist discourse? Specifically, we test the following hypotheses:

H1: Higher viewpoint diversity within a discussion thread is negatively associated with the diffusion of populist discourse.

H2: Higher levels of rationality within a discussion thread are negatively associated with the diffusion of populist discourse.

These hypotheses examine whether the intrinsic deliberative qualities of a discussion influence its subsequent diffusion, a question we explore using large-scale text analysis of Chinese online populism. This study will operationalize and empirically measure the deliberative concepts of “diversity” and “rationality” as they manifest in these discussions.

Method

Data Collection

Drawing on large‐scale discussions and interactions surrounding high-profile public incidents, Chinese scholars Cheng and Shi (2021) developed a coding scheme tailored to the Chinese context that identifies three recurrent populist issue types: anti-elite, anti-system, and national-populist themes. Anti-elite issues typically convey hostility toward groups with high socio-economic status, power-holding cadres, or intellectuals. Anti-system issues portray government performance in a negative light—highlighting, for instance, perceived judicial injustice, lack of credibility, or policy detachment from popular needs. National-populist issues manifest in two opposing attitudes toward the nation: uncritical patriotism on the one hand, and a deep sense of national inadequacy on the other (Cheng & Shi, 2021). Following this coding scheme, the present study operationalizes populism at the issue level and selects ten highly salient populist topics in contemporary Chinese online discourse (see Table 1).

We then employed a Python-based web scraper to collect all discussion threads and their corresponding posts related to these ten issues from the Zhihu platform. The data collection, completed in February 2022, yielded a dataset of 939 discussion threads containing a total of 212,218 posts.

Variable Measurement

Diversity

Viewpoint diversity in online discussions extends beyond simple partisan dichotomies or political polarization (Boukes, 2024; Camaj, 2021). It is more substantively characterized by the emergence of a broad spectrum of novel arguments and perspectives throughout the deliberative process (Ziegele et al., 2020). Informed by this conceptualization, this study operationalizes viewpoint diversity by combining topic modeling with an information entropy metric. The underlying rationale is that semantically similar arguments are grouped into distinct topics. Consequently, the distribution of text across these topics can serve as a robust indicator of the discussion’s heterogeneity.

Specifically, we employed the BERTopic model for topic extraction and calculated diversity using the normalized Shannon entropy formula:

The analysis was conducted at the issue level. For each distinct issue, all user posts from its corresponding discussion threads were aggregated into a single corpus. A separate BERTopic model was then trained on each of these corpora. This procedure yielded a normalized Shannon entropy score for each discussion thread (M = 0.43, SD = 0.40), quantifying its internal viewpoint diversity.

In this framework, entropy scores approaching 1 signify a more even distribution of discourse across multiple topics, thus indicating greater viewpoint diversity. Conversely, scores nearing 0 suggest that the discussion is concentrated on a limited number of topics, reflecting lower diversity.

Rationality

While rationality is a recognized cornerstone of deliberative quality, its operationalization presents a significant conceptual and methodological challenge. A primary difficulty stems from the considerable heterogeneity in coding schemes for rationality across studies (Camaj, 2021; Friess et al., 2021; Naab et al., 2025; Ziegele et al., 2020). Theoretically, a rational statement extends beyond a mere assertion to include evidence that is empirically verifiable or falsifiable (Habermas, 1984). Deliberation is thus conceived as a process of mutual critique of well-reasoned normative claims (Dahlberg, 2001).

Building on this definition, Rowe (2015) and Stromer-Galley (2007) delineated several interrelated dimensions of rationality. These require that participants: (1) express a clear opinion or position; (2) provide justification for their claims; and (3) support arguments with empirically verifiable evidence. Furthermore, both scholars emphasized the importance of topic relevance, positing that for deliberation to be effective, discussions must remain focused on the issue at hand and not deviate from the topic.

In a study of user comments during Facebook political debates, Camaj (2021) developed a content analysis scheme to operationalize these dimensions of rationality. This framework consists of four dimensions: (1) Opinion Expression, (2) Topic Relevance, (3) Justification, and (4) Evidentiary Support. Each dimension was coded dichotomously (1=present, 0=absent), and the sum of these scores formed the rationality index for each post.

Recognizing the robust construct validity and operational feasibility of this framework, the present study adopts it to assess the rationality of discussions on the Zhihu platform.

This study employed a large language model (LLM) to perform the content analysis. Recent research has shown that LLMs possess remarkable zero-shot learning capabilities for political text classification (Heseltine & Clemm von Hohenberg, 2024; Törnberg, 2024; Ziems et al., 2024), achieving high accuracy without requiring extensive labeled training data. For straightforward tasks like political ideology classification, their accuracy can exceed 90%.

As this research analyzes Chinese-language texts, a model proficient in the complexities of Chinese semantics was required. We therefore utilized Qwen-Max-Latest—a state-of-the-art LLM from Alibaba Cloud—as our annotation tool. The prompts provided to the LLM comprised three components: (1) a background description of the dataset, (2) specific task instructions, and (3) the rationality coding scheme (the full prompt and example outputs are available in Supplementary Appendix Table B1).

To validate the ’LLM’s coding performance, we randomly sampled 200 posts from the dataset. Two graduate students in journalism and communication were trained to manually code this sample according to the four rationality dimensions. Inter-coder reliability, measured using Cohen’s Kappa coefficient, exceeded 0.7 for all dimensions. The model’s classification performance is detailed in Table 2, achieving a macro-averaged F1 score of 0.87. This level of performance confirmed that the LLM’s annotations were sufficiently reliable for the subsequent analysis.

The LLM conducted post-level coding on the entire dataset of over 212,000 posts, with 90.5% of posts demonstrating at least one rationality dimension. The rationality of each discussion thread was then operationalized as the mean score of its constituent posts (M = 2.19, SD = 0.53).

Diffusion of Populism Discussions

The diffusion of each discussion was operationalized by two metrics: the number of votes (M = 7569.87, SD = 36876.29, Median = 46) and the number of comments (M = 1421.44, SD = 4825.25, Median = 42) for each thread.

Subsequently, employing the discussion thread as the unit of analysis, we constructed a regression model to test our hypotheses, with thread-level diversity and rationality as predictor variables and the number of votes and comments as outcome variables.

Results

Overall, the regression analyses for both votes and comments revealed significant fixed effects of the two deliberative quality dimensions (diversity and rationality) on the diffusion of populist discussions. Specifically, greater diversity was associated with an increase in both votes and comments, suggesting it facilitates the diffusion of populist discourse. Conversely, higher rationality was linked to a decrease in these engagement metrics, indicating it may curb the spread of such discussions. Furthermore, the significant random effects for issues underscore that the specific topic of a populist discussion also substantially influences its diffusion.

The data possess a clear hierarchical structure, as individual discussion threads (the unit of analysis) are nested within specific populist issues (see Supplementary Appendix Table A1 for sample sizes). This nested structure violates the independence assumption of standard regression, making multilevel modeling the appropriate analytical approach.

Furthermore, the outcome variables (votes and comments) are count data. Preliminary analysis revealed that both variables were substantially overdispersed, with their standard deviations far exceeding their means. Given the low proportion of zero values (13.63% for votes; 0% for comments), these characteristics warrant the use of multilevel negative binomial regression over a standard Poisson model.

Comments

To confirm the appropriateness of a multilevel approach for the comments variable, we first estimated an empty model with only a random intercept at the issue level. The resulting intraclass correlation coefficient (ICC) was 0.198, indicating that 19.8% of the variance in comment counts is attributable to between-issue differences. This substantial ICC value justifies the use of a multilevel model.

The full model, which included diversity and rationality as predictors, yielded a significant conditional overdispersion parameter (α = 3.92, p < .001). This confirms significant within-issue variance in comment counts and further validates our choice of a negative binomial regression model.

The fixed-effects results (Table 3) show that diversity had a significant positive effect on the number of comments (B = 4.16, p < .001). This indicates that a 0.1-unit increase in diversity is associated with a 0.416 increase in the log-expected count of comments, which corresponds to an approximate 52% increase in the expected number of comments. This finding suggests that greater viewpoint diversity in online populist discussions in China stimulates user engagement and amplifies the discussion’s reach.

In contrast, rationality was negatively associated with comment volume (B = -0.85, p = .003; see Table 3). This coefficient suggests that a one-unit increase in the rationality score is linked to an approximate 57% decrease in the expected number of comments. In essence, when populist discussions feature more logical and structured argumentation, they attract less user engagement in the form of comments, thereby constraining their diffusion.

Votes

An initial empty model for the votes variable yielded an ICC of 0.232 and a conditional overdispersion parameter (α) of 6.42. These findings confirmed that a multilevel negative binomial regression was also the appropriate analytical approach for this outcome.

Given the presence of zero values in the votes variable, we conducted further analyses to test for potential structural zero-inflation. Specifically, we compared the model fit of the multilevel negative binomial regression model against its zero-inflated counterpart (ZINB). The standard negative binomial model demonstrated a superior fit, yielding lower Akaike Information Criterion (AIC = 13,113.65) and Bayesian Information Criterion (BIC = 13,137.87) scores than the ZINB model (AIC = 13,115.65; BIC = 13,144.72). Furthermore, diagnostics performed with the DHARMa package in R confirmed this finding, showing that the observed frequency of zeros was significantly lower than that predicted by the model (p < .001). Collectively, these results indicated an absence of substantial zero-inflation and confirmed that incorporating a zero-inflation component did not improve model parsimony or performance. Therefore, the more parsimonious multilevel negative binomial regression model was employed for all subsequent analyses.

The regression analysis revealed that the influences of diversity and rationality on votes paralleled their effects on comments. More specifically, diversity was a significant positive predictor of votes (B = 5.57, p < .001), whereas rationality exhibited a significant negative association (B = -1.16, p = .004; Table 4). Echoing the pattern observed for comments, these findings indicate that a greater diversity of viewpoints promotes the diffusion of populist discussions by increasing voting engagement. Conversely, higher rationality serves to inhibit this diffusion.

In summary, the results failed to support Hypothesis 1 (H1) but supported Hypothesis 2 (H2).

Notably, significant random intercept variances were observed at the issue level in both regression models: 1.51 for comments (Table 3) and 1.67 for votes (Table 4). This finding underscores that substantial variation in discussion diffusion is attributable to the specific populist issues themselves. In other words, after accounting for predictors like diversity and rationality, the topic of discussion remains a powerful, independent driver of its diffusion.

Therefore, the spread of populism on Chinese social media is driven not only by the quality of deliberation but also, critically, by the substance of the issues under discussion.

Discussion

This study examined how two core dimensions of deliberative quality—viewpoint diversity and rationality—relate to the diffusion of populist discussions on Chinese social media. Across two engagement metrics (comments, votes), we find a consistent dichotomy: greater diversity amplifies diffusion, whereas greater rationality dampens it. In addition, sizeable issue-level random effects indicate that diffusion varies systematically by topic, even after accounting for deliberative quality.

The positive effect of diversity aligns with evidence that exposure to differing perspectives heightens user participation in discursive spaces (Morey, Kleinman, & Boukes, 2018) and with a program of work showing that perceived controversy boosts engagement intentions (Ziegele et al., 2014, 2018, 2020). A parsimonious mechanism is that heterogeneity introduces contestable claims that challenge prior beliefs, thereby increasing attention and affective involvement and prompting expressive action (see also Diakopoulos & Naaman, 2011; Shoemaker, 1996; Singer, 2009). In short, diversity operates as a communication amplifier in populist threads.

By contrast, higher rationality—more reason-giving, sourcing, and argumentation—is negatively associated with diffusion. This pattern complicates deliberative ideals (Habermas, 1996) and coheres with findings that, in affect-laden and polarized contexts, emotional and confrontational cues outperform reasoned argument in eliciting interaction (Camaj, 2021; Ziegele et al., 2020). A plausible mechanism is a cooling effect: rational posts lower emotional arousal, raise cognitive load, disrupt outrage/reinforcement cycles, and weaken bandwagon signals that platforms and users often reward—thereby curbing participatory cascades even as arguments improve.

Finally, the persistence of topic-specific variance underscores that populist diffusion is not only about how people deliberate but also what they deliberate about. Issue framing and emotional resonance appear to condition the returns to diversity and rationality, consistent with views of populism as an issue-sensitive communication paradigm (Waisbord, 2018). Theoretically, our results bridge deliberative democracy and populism by showing that diversity can act as a double-edged amplifier in populist contexts, whereas rationality functions as a diffusion brake. Practically, they caution against diversity-only interventions and support community/design measures that elevate reason-giving (e.g., prompts for evidence or sources) while accounting for topic-level virality in governance and platform curation.

From a practical standpoint, the findings offer empirical insights into the diffusion mechanisms of populist discussions in the Chinese online sphere, providing an evidence-based foundation for platform governance. On the one hand, the findings indicate that greater diversity does not necessarily inhibit the diffusion of populist issues; only under the condition of fostering rational deliberation can online populism and public opinion crises be effectively contained. On the other hand, the very nature of the issue significantly shapes dissemination outcomes, as certain topics are inherently more likely to trigger interaction. This offers a new perspective for platform governance. In practice, platforms may adopt differentiated review or tiered management strategies based on issue characteristics, strengthening manual review for topics prone to rapid spread, while simultaneously adjusting algorithmic design to enhance the recognition and visibility of rational expressions. In addition, social platforms themselves can optimize interface design to encourage users to provide more reasons and evidence when participating in discussions, thereby increasing the overall rationality of deliberation.

In summary, this study not only furnishes empirical evidence on the complex effects of deliberative quality on the diffusion of populist discourse but also makes key theoretical advances. It does so by contextualizing deliberative democratic theory for new digital environments, deepening the understanding of digital communication structures, and refining the mechanisms of populist propagation.

Limitation and Future Study

First, in terms of variables, this study did not explicitly examine how issue characteristics shape the diffusion of populist issues. For instance, classifying issues by thematic content (e.g., anti-elitism, anti-establishment sentiment, or national populism) could yield greater explanatory power. Second, the data source is relatively limited, focusing exclusively on Zhihu. As China’s largest knowledge-oriented Q&A community, Zhihu exhibits distinct elitist features in both platform culture and user composition: its users are predominantly well-educated members of the middle class, with a concentration among younger cohorts (74% under the age of 30); moreover, the platform often hosts in-depth discussions or debates on socio-political issues (Peng, 2020; Zhu, 2024). This means that the data used in this study reflect only how groups with relatively strong discursive power in China construct populist issues and thus face clear limitations in representing the grassroots ecology of “bottom-up” online populism.

In addition, because online information flows are generally subject to supervision and regulation, the data collected in this study are inevitably filtered through China’s internet censorship mechanisms. Yet this dilemma is hardly unique to this research, as it is a common challenge for social media text studies. Recent scholarship has pointed out that China is not the most heavily censored country worldwide (Al-Zaman, 2025), and that Chinese social media still contains a large amount of open, sustained, and critical discussion of social issues that are officially recognized and do not touch upon the political core (Rauchfleisch & Schäfer, 2015). To minimize bias stemming from censorship, this study selected social issues that had previously triggered widespread and enduring online debate in China, thereby ensuring sufficient visibility of related discussions.

Finally, with respect to the measurement of the core variable—rationality—this study employed large language model (LLM)-based coding. While the model’s performance was satisfactory, its capacity to identify non-traditional forms of rational expression that frequently appear on Chinese social media—such as metaphor, allusion, and irony (Wu & Fitzgerald, 2021)—remains untested, raising the possibility that the rationality present in such discourses may have been underestimated.

Future research could first consider refining models of how diversity and rationality influence the diffusion of populist issues by incorporating issue characteristics as a moderating variable. This would allow for a more systematic examination of differences in the online dissemination potential of various types of populist issues. In addition, future studies should move beyond the limitation of a single platform, collecting data from as many platforms as possible, or employing cross-platform comparisons that extend the analysis to widely used and more grassroots-oriented social media such as Weibo and Douyin, in order to fully capture the interaction between online deliberation and the spread of populist discourse in China. Methodologically, researchers may further interpret cases misclassified by large language models, testing whether the algorithm exhibits systematic bias in detecting non-traditional forms of rational expression. Based on such evaluations, future work could optimize prompt design or adopt few-shot learning approaches to further enhance the validity and robustness of LLM-based coding.

—-

(*) School of Journalism and Communication, Tsinghua University, Beijing, China. Email: suy21@mails.tsinghua.edu.cnORCID:0009-0008-8874-611X

(**) Corresponding author: Tongtong Li, 400 Guoding Road, Yangpu District, Shanghai 200433, China. Email: litt23@m.fudan.edu.cn, School of Journalism, Fudan University, Shanghai, China. ORCID:0009-0003-2895-2716  

 —-

Declaration of Interest Statement

No potential conflict of interest was reported by the author(s). 

 —-

Funding

The author(s) reported there is no funding associated with the work featured in this article.

Footnotes

1. The “Driving a Mercedes into the Forbidden City” incident was sparked on January 17, 2020, when a Sina Weibo user posted several photos of herself and a friend posing with a luxury car inside the Palace Museum—China’s most iconic imperial palace and a UNESCO World Heritage site. In her post, she wrote, “On Monday, the Palace Museum was closed, so I hurried over, hid from the crowds, and went to play in the Palace Museum.” Since 2013, private vehicles have been strictly prohibited from entering the museum grounds, which are typically closed for maintenance on Mondays.

2. The Xinjiang Cotton Incident refers to a wave of public backlash in China that erupted in March 2021, after several international brands—including H&M, Nike, and Adidas—announced they would stop sourcing cotton from Xinjiang due to allegations of forced labor and human rights abuses in the region.