Abstract
Drawing on Moral Foundations Theory and recent network diffusion studies, we examine the toxicity spread in US immigration-related discourse on X (N = 305,878 tweets, January 2021 to April 2024). Using AI-assisted content analysis, this study examines how psychological predispositions (issue stance and moral foundations) and network positions predict users’ toxicity roles (i.e. amplifiers, copycats, and attenuators). Multinomial regression and random forest models reveal that amplifiers are typically centrally embedded users who score high on binding moral values, especially sanctity. Attenuators often hold pro-immigration stances and receive more inbound attention, suggesting reputational restraint and network moderation. Copycats, situated in dense, reciprocated ego networks, tend to reciprocate and mimic the tone of their peers. Contrary to expectations, amplification is not driven by bridge actors but by highly central users. This typology enhances our understanding of how moral convictions and network structures influence propagation, with implications for platform design and network intervention strategies.
Introduction
In an era where social media has become a primary venue for political discourse, concerns about the pervasive incivility have intensified (Kushin and Yamamoto, 2025; Lu et al., 2023; Ng et al., 2020). Incivility ranges from insults and deception to efforts that undermine democratic deliberation (Bentivegna and Rega, 2022; Rowe, 2015). Incivility diffusion is particularly troubling in participatory platforms, where users are not passive consumers but active participants who amplify, adapt, and recirculate incivility (Hmielowski et al., 2014; Vaidya et al., 2024).
For communication scholars, understanding incivility is essential for its prevalence and consequences for democratic discourse (Sobieraj and Berry, 2011). Incivility has documented negative effects, including eroding trust, reducing news engagement, and heightening intolerance (Lu et al., 2023; Novotná et al., 2023; Van’t Riet and Van Stekelenburg, 2022). Developing an integrated understanding of incivility’s antecedents and effects is critical for fostering healthier discursive environments (Ng et al., 2020). This current research focuses specifically on a subtype of incivility—toxicity—which operationalizes individual-level impoliteness and delegitimizing expressions (Bentivegna and Rega, 2022) as explicitly hostile incivility, including personal attacks, insults, obscene language, and aggression. Toxicity is used to refer to this specific individual-level incivility (not institutional-targeting incivility) that this study focuses on, and incivility will be mentioned when the broader theoretical backdrop is mentioned.
Despite growing interest in toxicity, communication research has only recently begun to explore how uncivil content diffuses and shapes discursive norms. Hmielowski et al. (2014) offer an insightful perspective by conceptualizing toxicity diffusion as a socialization process: repeated exposure to online political discussions can normalize flaming, especially among individuals predisposed to verbal aggression. In this view, toxicity does not merely spread through visibility but gains legitimacy over time, encouraging users to adopt more aggressive communication styles.
Moreover, to understand individuals’ roles in the spread of toxicity on social media, Vaidya et al. (2024) identify three user roles—amplifiers, copycats, and attenuators—based on how they spread toxicity. Amplifiers spread more than they receive, attenuators spread less, and copycats spread about the same. Their framework shows that individual traits and network positions influence toxicity diffusion. The current study examines how issue stance, moral values, and network placement predict users’ roles in spreading toxicity. Vaidya et al.’s framework highlights the various ways users contribute to toxicity and suggests that diffusion is mediated by both individual predispositions and structural positions within communication networks. Drawing on these, the present study extends previous research and identifies what individual predispositions (e.g. issue stance and moral values) and structural positions (e.g. different network positions) predict users’ roles in toxicity diffusion.
By systematically examining the antecedents of these roles, our study helps to complete the puzzle of toxicity diffusion. Rather than treating amplifiers, copycats, and attenuators as static categories, we reveal the complex interplay of psychological orientation and networked context that drives individuals to assume each role. This integrated perspective provides a more comprehensive understanding of how toxicity takes root and spreads online.
In X (our research context), Musk’s 2022 acquisition reduced the trust and safety workforce by 80%, advocating for free speech, while hate speech surged 50% (Brewster, 2024; Hickey et al., 2025), exemplifying tensions between free expression and user safety (Schoenebeck and Blackwell, 2021). Our findings reveal that toxicity spread is shaped by users’ moral expressions and network positions, suggesting that effective platform governance must move beyond reactive, content-level moderation to understand how psychological orientations and relational dynamics enable or constrain toxic discourse propagation.
We situate our research within the context of the US immigration debate during the Biden administration, as immigration has long been an issue marked by controversial policies, intensified polarization, and elevated online engagement. We begin by reviewing recent communication research on toxicity, with emphasis on how toxicity circulates on social media. Anchored in emerging models of toxicity diffusion, we focus on the roles of amplifiers, copycats, and attenuators as key actors in shaping discursive trajectories. We then examine individual predispositions—specifically, users’ immigration stance and moral worldviews, as inferred from their language—and contextual network factors that may predict these roles.
The diffusion of toxicity on networked social media
Incivility and toxicity research in communication
Communication scholars have extensively examined incivility, recognizing it as a pervasive and consequential element of contemporary political discourse (Bentivegna and Rega, 2022; Muddiman and Stroud, 2017; Rossini, 2022; Rowe, 2015). The definition of political incivility remains contested. Some argue for a narrow definition that describes incivility as behaviors that threaten democracy, deny individuals’ freedoms, and stereotype social groups, distinguishing it from politeness in personal exchanges (Papacharissi, 2004; Rowe, 2015). However, more recently, scholars have explored a broader conceptualization of incivility that incorporates both interpersonal impoliteness and institutional norm violation. This scholarship has moved beyond framing incivility merely as a breakdown of politeness norms, instead reconceptualizing it as a multifaceted phenomenon with diverse discursive forms and social consequences. For example, Bentivegna and Rega (2022) propose a three-dimensional model of political incivility comprising impoliteness, individual delegitimization, and institutional delegitimization.
This study captures specifically a subdimension of incivility at the personal level, including impoliteness and individual delegitimization in immigration-related expressions. Impoliteness in political expressions, as Bentivegna and Rega (2022) described, is the most common form of incivility, encompassing behaviors such as insulting, making fun of, belittling, and attacking. Individual delegitimization includes indicators such as the “use of racial, sexist, ethnic, and religious slurs,” “the use of discriminatory stereotypes,” “threatening harm,” “excluding the other,” and so on (p. 5).
Specifically, we adopt toxicity as the manifestation of interpersonal incivility and individual delegitimization in expressions, capturing explicit attacks, insults, threats, profanity, identity attacks, and aggressive language that cross the boundaries of acceptable human exchange and delegitimize others through “harmful exclusion and discrimination” (Wulczyn et al., 2017: 5). Institutional delegitimization targets democracy for the purpose of impeding decision-making or legislation (Bentivegna and Rega, 2022). Although possibly violating public norms, institutional delegitimization is a subtler rhetoric that may sound like a calm and logical critique of the system. We argue that institutional delegitimization spreads through mechanisms distinct from those of interpersonal toxicity. For example, interpersonal toxicity is often driven by affective reactions that trigger network cascades, whereas institutional delegitimization spreads more through ideological alignment and persuasion, which are out of scope for this research. By focusing on toxicity, we inquire how it cascades through network mechanisms.
The narrow focus on toxicity is supported by precedents in communication research. A substantial body of research has focused on the effects of individual-level toxicity on citizens, particularly in terms of political trust and participation (Boukes, 2025). Other studies have investigated the nature and content of toxicity, including extreme expressions such as outrage discourse, which seeks to provoke emotional responses (Sobieraj and Berry, 2011), as well as the dynamics of toxicity in specific contexts like pandemic-related Facebook discussions (Novotná et al., 2023) or non-partisan comment threads (Lu et al., 2023). More recently, scholars have explored interventions aimed at reducing toxicity, such as self-regulation strategies, which show promise in curbing uncivil behavior in online settings (Kushin and Yamamoto, 2025).
In addition to the well-documented negative consequences of incivility for deliberative norms and discourse quality, a growing body of scholarship highlights its potentially functional and strategic dimensions. This perspective helps explain why users do not simply avoid incivility but may actively produce, engage with, and amplify it. Research shows that the diffusion of incivility is closely intertwined with political partisanship and affective polarization. In their analysis of 9.6 million comments on The New York Times, Muddiman and Stroud (2017) demonstrate that partisan incivility tends to increase user engagement, even while simultaneously attracting abuse flags. This paradox underscores the dual nature of incivility: it may be normatively disapproved of at an abstract level, yet behaviorally rewarded through attention, interaction, and amplification. Rather than passively tolerating uncivil content, users often engage with and promote uncivil messages that align with their partisan identities. In this sense, incivility can function as a form of strategic identity signaling—mobilizing in-group solidarity while demarcating out-groups.
Similarly, Rega et al. (2023) show that political elites may strategically deploy uncivil rhetoric, particularly on contentious issues such as immigration, to energize audiences, stimulate discussion, and shape the tone of subsequent user interactions. Such findings suggest that incivility can cascade through networks not merely as spontaneous norm violations but as a socially and politically consequential communicative strategy. More broadly, incivility may strengthen perceived group cohesion and boundary maintenance among ideologically aligned users (e.g. Frimer et al., 2023; Heseltine and Dorsey, 2022).
Taken together, this line of research positions incivility as a multifaceted, context-dependent phenomenon that can be strategic, functional, and engagement-rewarded within polarized environments. It is shaped by platform affordances (Schulze et al., 2024), user identities, issue domains, ideological alignments, and broader political dynamics. Against this backdrop, the present study examines the diffusion of individual-level toxicity as a networked communication process, focusing on the distinct structural and psychological roles users play in shaping its trajectory. By situating our analysis within both the normative and strategic literatures, we provide a more comprehensive account of why uncivil content persists and spreads online.
Diffusion of toxicity in communication networks
The diffusion of toxic expressions is a networked process shaped by users’ positions within digital networks and platform affordances (Lee et al., 2019; Schulze et al., 2024; Szabó et al., 2021). Toxic expressions cluster in threads and polarized subcommunities, indicating contagion-like effects where network structure either amplifies or constrains hostile discourse. While network structures play an important role, the spread of toxicity is also fundamentally driven by individual-level agency and communicative behaviors. Hmielowski et al. (2014) argue that online political discussion itself can socialize users into accepting flaming—an extreme form of toxic expression—as a normative mode of interaction. Through repeated exposure to hostile exchanges in digital forums, individuals may come to view such behavior as legitimate, particularly if they already exhibit high verbal aggression. This internalization of norms subsequently increases their intention to flame. Importantly, Hmielowski et al. show that the opportunity to engage in such discourse—afforded by participatory platforms—is a key enabling condition. Toxicity, then, is not a passive outcome of structural conditions; it is actively produced and reinforced by users who adopt and reproduce aggressive styles in their everyday political talk. When users repeatedly participate in contentious discussions and encounter minimal sanctions or even social rewards for toxicity, the behavior becomes routinized. Thus, the diffusion of toxic clusters in digital networks is best understood as a dual process, emerging from the interplay between individual communicative tendencies and the structural affordances of the online spaces they inhabit.
Building on our conceptualization of toxicity diffusion as a networked communication process, we draw further on Vaidya et al.’s (2024) user typology to illuminate the distinct roles individuals may play in propagating toxicity.
Toxicity spreader typology
To understand the network diffusion of toxicity, Vaidya et al. (2024) proposed a typology that classifies users based on their role in spreading toxicity: amplifiers, copycats, and attenuators. This typology draws on a networked perspective, comparing the toxicity users receive (in incoming messages) with the toxicity they generate (in outgoing messages). Amplifiers are users who send more toxic messages than they receive, thereby intensifying toxicity within the network. Attenuators, in contrast, respond to toxicity with more civil discourse, thereby mitigating its spread. Copycats fall in the middle, matching the toxicity they receive and reinforcing prevailing communicative norms.
This approach moves beyond merely identifying toxic content to detecting the dynamic roles users play in its diffusion. It echoes calls from scholars like Chen et al. (2024), who emphasize the need to analytically separate exposure to and expression of toxicity. Moreover, each user role has conceptual parallels in prior literature. Amplifiers resemble flamers and trolls, who spread emotionally charged or provocative content to elicit strong reactions (Phadke and Mitra, 2021). These users often drive high engagement and are commonly associated with misinformation. Copycats, by mirroring the toxicity in their environment, reinforce group norms and emotional climates—a dynamic resembling emotional contagion and the “clinching effect” described by Brundidge and Garrett (2024). Attenuators, though less studied, play a critical yet often invisible role (Cook et al., 2021), acting as buffers that de-escalate conflict and reduce the spread of hostile sentiment.
Despite the promise of this typology, little is known about the factors that drive individuals to assume different roles in the diffusion of toxicity. This raises a central question: what drives users to assume roles of amplifiers, copycats, or attenuators? To address this question, we explore both users’ psychological dispositions and structural positions within communication networks as potential key influences on their behavior, as further discussed below.
User psychological dispositions for toxicity spreaders
Users’ issue stance and toxicity predisposition
Political toxicity is not merely a breach of communication norms—it also serves as a tool for emotional expression, identity signaling, and ideological alignment (Rossini, 2022; Rowe, 2015). Recent scholarship has shown that individuals with differing political or issue positions may both express and interpret toxicity in distinct ways. For instance, Kosmidis and Theocharis (2020) find that toxic commentary can elicit positive emotional responses among ideologically aligned audiences, suggesting that toxicity may function less as a source of alienation and more as a form of tribal reinforcement.
Immigration is a historically divisive and emotionally charged issue. Online discourse around immigration often reflects profound ideological divides, and users with different stances tend to engage in distinct communicative behaviors (Nortio et al., 2021). Anti-immigration messaging frequently relies on fear-based narratives, dehumanizing language, and moral condemnation—forms of toxicity that are designed to provoke, polarize, and mobilize (Farjam and Segesten, 2024). Such messaging aligns with the amplifier role in the typology of toxicity diffusion: users who produce more toxic content than they receive, thereby escalating emotional conflict and reinforcing group boundaries.
In contrast, pro-immigration discourse typically emphasizes values such as inclusion, empathy, and humanitarianism (Stravato, 2024). Although pro-immigration users may encounter uncivil content, they may not reciprocate in kind. This restraint reflects the behavior of attenuators, who are exposed to toxicity but do not perpetuate it. Meanwhile, users with neutral or ambivalent views on immigration may lack strong ideological convictions and instead adopt the tone of the surrounding discourse. These users resemble copycats, who amplify or attenuate toxicity based on the communicative environment rather than internalized attitudes.
Together, these patterns suggest that a user’s issue stance may predispose them toward a particular role in the diffusion of toxicity. Anti-immigration users, motivated by ideological intensity, are more likely to function as amplifiers. Pro-immigration users, guided by empathic concerns, are more likely to be attenuators. Neutral users, lacking a clear ideological direction, may be susceptible to social mimicry and more likely to behave as copycats. Therefore:
Users’ moral foundation and toxicity predisposition
Moral Foundations Theory (MFT) distinguishes between individualizing foundations (care and fairness) that prioritize individual rights and binding foundations (loyalty, authority, and sanctity) that emphasize group cohesion (Haidt and Graham, 2007). These foundational differences may play a critical role in shaping users’ predispositions toward toxicity. Prior work suggests that moral convictions are not only affectively charged but also socially consequential—they motivate defense of in-group values and can justify harsh rhetoric against perceived violators (Day et al., 2014). Binding moral foundations have been linked to a stronger motivation to defend group norms and to condemn out-groups. For example, Baldner and Pierro (2019) found that individuals with a higher need for cognitive closure were more likely to adopt anti-immigrant attitudes, and this relationship was mediated by binding foundations. This suggests that loyalty, authority, and purity may facilitate group-focused moralization that can manifest as prejudice or antagonism toward outsiders such as immigrants.
In contrast, individualizing moral foundations is more attuned to minimizing harm and promoting fairness across group lines. These values tend to encourage inclusive and empathetic engagement, potentially serving as a buffer against toxicity. Supporting this distinction, Kim et al. (2022) found that in the context of anti-Asian hate speech during the COVID-19 pandemic, tweets containing care- and fairness-based moral language were associated with more civil expressions, whereas tweets invoking group-based threats were more likely to contain toxicity.
The psychological mechanisms linking moral foundations and toxicity may also operate through emotional and cognitive reactions to online disagreement. Hwang et al. (2018) demonstrated that exposure to uncivil discourse can trigger negative emotions, such as moral indignation, which subsequently leads to increased closed-mindedness and toxicity. If binding moral foundations heighten sensitivity to perceived violations of group norms, they may amplify this reactive cycle, fostering aggressive rhetoric in defense of group identity. Conversely, individualizing foundations may temper such reactions through appeals to shared humanity and justice. Building on this theoretical foundation, we propose the following hypothesis:
Structural conditions for toxicity spreaders
In online discourse, the structural position of users in a network shapes not only what information they are exposed to but also how they behave—particularly when it comes to expressing toxicity (Hopp and Vargo, 2019). Network positions can serve as filters or amplifiers of discourse norms, and users’ exposure to norm-violating content may condition their own communicative behaviors. A growing body of work suggests that toxicity is not just a matter of individual disposition but also of external context, visibility of interpersonal norms, and the embeddedness of communicator networks (Bormann et al., 2022; Rains et al., 2017).
Users’ structural positions shape not only what they are exposed to but also how frequently their moral orientations are activated in contentious exchanges. In this article, we argue that network location could condition when and how these moral motivations translate into behavior. Specifically, users occupying bridging positions—those connecting otherwise disparate communities—are particularly susceptible to heterogeneous and ideologically cross-cutting exposure (Hopp and Vargo, 2019). Operationalized through measures such as betweenness centrality, bridging positions increase contact with conflicting moral narratives and identity claims. Such encounters can heighten moral salience, especially when group loyalty, authority defense, or sacred values appear threatened. For users high in binding foundations, exposure to cross-cutting disagreement may therefore activate defensive or boundary-reinforcing responses, increasing the likelihood of toxic amplification. Although bridging capital can confer cognitive benefits, it also generates psychological strain; cross-cutting exposure has been shown to intensify affective polarization under high-toxicity conditions (Lin et al., 2023), and exposure to toxic discourse predicts subsequent toxic expression (Kim et al., 2021). From a joint MFT-network perspective, bridge-positioned users are structurally exposed to morally charged conflict, and those with binding moral orientations may be especially likely to translate that exposure into amplification. Thus, amplification is not merely a positional artifact but the product of moral motivation activated within structurally conflict-prone locations (Lee et al., 2019).
In contrast, highly central users—those with substantial influence and visibility, captured by eigenvector centrality, PageRank, or degree centrality—occupy positions characterized by broad reach and reputational accountability. Centrality increases audience scope and status visibility, making norm violations more consequential. For users high in individualizing foundations, such visibility may reinforce commitments to fairness and harm reduction, strengthening their role as attenuators of toxicity. Even for users with stronger binding orientations, reputational incentives and cross-audience accountability may constrain overtly toxic expression. As Rains et al. (2017) argue, visible identity performance promotes socially confirming behavior. In this way, central positions provide structural incentives that align particularly well with individualizing moral motivations, increasing the likelihood of toxicity attenuation.
Finally, users embedded in dense, close-knit communities—characterized by high local clustering, reciprocity, and ego network density—operate within tight feedback loops where conformity pressures are pronounced. In such environments, moral expression is filtered through local norms. If a subgroup’s dominant tone frames toxicity as a signal of loyalty or identity defense, users high in binding foundations may be especially prone to mimic and reproduce that tone. Conversely, if civility is normatively rewarded, users high in individualizing foundations may reinforce those standards. Prior research shows that communication norms can suppress toxicity when environments reward civility (Bormann et al., 2022), while toxicity can function as a performative identity signal (Rains et al., 2017). Thus, densely embedded users may act as “copycats” of toxicity—not because of independent moral recalibration but because moral expression becomes locally norm-bound within cohesive clusters.
Taken together, network positions structure exposure, accountability, and conformity pressures, while moral foundations shape how individuals interpret and respond to these structural conditions. Toxicity diffusion, therefore, reflects the interaction between moral motivation and structural opportunity. Accordingly:
While individual predispositions are critical for understanding users’ psychological motivations for spreading or suppressing toxicity, they are only part of the equation. Online behaviors unfold within dynamic social contexts, where users are embedded in networks that shape what content they see, engage with, and emulate. Prior research suggests that both dispositional traits and structural factors—such as users’ connectivity, exposure to toxicity, and centrality in the network—jointly influence diffusion behaviors (Bakshy et al., 2011; Zhang et al., 2018). Yet, little is known about the relative importance of these different types of predictors in shaping users’ roles in the propagation of toxicity. To address this gap, we ask:
Method
Sample
Tweets collection
Data were collected from X using Apify. The data spans from 20 January 2021, when President Biden assumed office, to April 2024, when we started data collection. We carefully curated a list of relevant and top-ranked keywords for immigration-related data collection (see Appendix A, Supplementary Materials for detailed procedures). To distinguish tweets about US immigration from those about immigration elsewhere, we employed a hybrid approach combining manual coding and automated classification (Supplementary Appendix B). Four coders annotated 3654 posts with high agreement (κ = .90), which were used to further train and validate a DistilBERT classifier (89% accuracy) that was applied to all posts, retaining 305,878 posts out of all 413,326 tweets that were US-related.
Measures
Toxicity
The Detoxify model (Hanu and Unitary team, 2020), a BERT-based model trained on Wikipedia comments (Wulczyn et al., 2017), demonstrated high performance across contexts and platforms, including Twitter/X (Bouchaud et al., 2023; Farjam and Segesten, 2024). To validate performance in our immigration discourse context, two human coders evaluated the model’s validity and reliability. See Supplementary Appendix C for details.
Users’ immigration stances
Stance detection was performed using a combination of manual and automated processes. Four human coders manually labeled posts into pro-, anti-, or neutral immigration stances (N = 2829, κ = .81). In addition, GPT-4 was used for scaled classification (N = 8447, accuracy = 0.98 based on 1000 overlapping samples with human coders). Finally, using both human- and AI-labeled data (N = 11,276), a step-wise distilBERT model was trained, reaching a satisfactory accuracy of 0.81 and 0.88, and was applied to the rest of the data. See Supplementary Appendix D for details.
Toxicity spreader typology
We adopted Vaidya et al.’s (2024) toxicity role classification method. Users’ average toxicity of their own post was calculated (Tuser), and the toxicity of their neighbors’ incoming tweets was also calculated (Tneigh). Then, the differences between the scores were measured as (ΔT = Tuser−Tneigh). Toxicity role classification was based on the distribution of ΔT across all users in the community. See Supplementary Appendix E.
Morality
We adopted MFormer (Nguyen et al., 2024), a RoBERTa-based transformer model trained to detect the five dimensions of the moral foundation in each tweet: care, fairness, loyalty, authority, and sanctity (Haidt and Graham, 2007). MFormer provides probability scores between 0 and 1 for each of the five dimensions to indicate the likelihood of each moral dimension being emphasized in the text. The MFormer model was trained on 87,135 annotated texts from Twitter, news articles, and Reddit, with a range of different political and social topics, maintaining a high model performance across contexts, achieving a 4–12% improvement over the widely adopted lexicon-based methods (e.g. Frimer et al., 2019; Graham and Haidt, 2012; Hopp et al., 2021; c.f. Nguyen et al., 2024). We used MFormer to label each tweet, and then, for each user, we calculated an average morality score based on all their posts.
Network metrics
First, degree centrality was measured using normalized in-degree and out-degree values, reflecting users’ incoming and outgoing connections. We also captured eigenvector centrality, reflecting how connected people are to central others. We also estimated page rank, another centrality measure that is used to identify how important or influential a node is. It is the probability of a random walk to each node, taking into account the influential nodes to which it is linked. Second, betweenness centrality quantified how often a user lies on the shortest paths between other nodes, indicating their role as a broker within the network; this measure was also normalized. Ego density measures the proportion of actual ties among a node’s neighbors relative to all possible ties, also showing how constrained or locally supported they are in their local networks. Third, reciprocity measures the number of mutual ties out of the total number of directed edges a person is connected to, reflecting the mutuality of each person’s local network. Fourth, closeness centrality was calculated as the average shortest path from a user to all other nodes and normalized to enable comparison. Fifth, we computed local transitivity, defined as the proportion of a user’s neighbors that are also connected to one another. This is basically an attention-weighted structural authority in information networks. All metrics were computed using the igraph package in R.
Control variables
In addition, we controlled for user status on X. These included retweet count, reply count, like count, and quote count—metrics that capture content virality and user interaction. In addition, author-level features such as follower count, following count, and account tenure were included to represent users’ social reach and platform experience.
Network construction
We constructed a directed mention network where nodes are Twitter users and directed edges represent mention interactions. Specifically, when a user tweets content containing another user’s handle (e.g. “@username”), we create a directed edge from the sender to the mentioned user. We focus on mentions because they represent “actor-motivated communication” with a drive to create conversations (Yang and Saffer, 2021: 2910), distinguishing them from reposts that are content-driven (Yang and Saffer, 2021). This network structure allows us to examine how users are positioned in immigration-related conversations and original content creation and conversational interactions, rather than through passive reposting behavior, where toxicity may not necessarily reflect the user’s own expressions or targeting intent. While the retweet network reflects content amplification, ideological alignment, and endorsement (Barberá, 2015), it does not represent users’ original toxicity level and language features. Practically, we chose mention over retweet because data collection from Apify does not yield retweet data.
Analytic procedure
We conducted a series of multinomial logistic regressions. Prior to modeling, we standardized all independent variables using z-score transformation via scikit-learn, ensuring comparability across features and enhancing model stability. Variance inflation factor (VIF) tests were run and showed no multicollinearity. Given the imbalance in the dependent variable’s class distribution, we computed class weights using the compute_class_weight function. This approach compensates for unequal class representation by assigning higher weights to underrepresented categories, thereby reducing bias. Next, the multinomial logistic regression models were trained using the lbfgs solver, incorporating the computed class weights to further address imbalance. Models were fit on the training data with a maximum of 1000 iterations to ensure convergence. We then evaluated performance on the test set using metrics such as classification accuracy, confusion matrix, and precision-recall scores. In addition, to assess which predictors are most influential, we built a random forest classifier—an ensemble, nonparametric method well suited for modeling complex, nonlinear relationships and variable interactions.
Results
The H1 predicts that anti-immigration users are more likely to act as amplifiers, pro-immigration users as attenuators, and neutral users as copycats. Although the overall classification accuracy is modest (precision = .81, accuracy = .57), the direction of the coefficients for immigration stance supports H1. That is, anti-immigration users are more likely to serve as amplifiers of toxicity, pro-immigration users as attenuators, and neutral users as copycats (see Table 1, Model 1). More precisely, a stronger pro-immigration stance was associated with lower odds of being an amplifier and higher odds of being an attenuator, relative to being a copycat.
Multinomial Logistic Regression Coefficients Predicting Roles In Toxicity Diffusion.
We Are Using Scikit-Learn’s Logistic Regression Model, Which Is A Softmax-Based Approach. There Is No Specific Reference Category. Positive Coefficients Indicate Increased Odds Of Being In The Target Class Compared To The Reference.
The H2 proposes that users with stronger individualizing foundations (care and fairness) are more likely to serve as attenuators, while those with stronger binding foundations (loyalty, authority, and sanctity) are more likely to serve as amplifiers. The model shows better accuracy than Model 1 (precision = .84, accuracy = .68) and offers partial support for H2. Specifically (Model 2), binding moral foundations show the most robust associations: sanctity has a strong positive coefficient for the amplifier role (b = 0.75) and a negative coefficient for attenuators (b = −0.20), supporting H2’s prediction that sanctity is linked to amplification of toxicity, consistent with the prediction direction for H2. Loyalty and authority are weakly associated with the amplifier role (b = −0.03 and b = 0.001, respectively), suggesting partial support in line with H2. In addition, authority also shows a weak positive association with being an attenuator (b = 0.005), reflecting that people with high authority tend to play active roles (either attenuating or amplifying toxicity) instead of copying their neighbors’ toxicity, although the effect is small. In other words, people with high loyalty tend not to be in an amplifier role, but people with high authority slightly increase the odds of being an amplifier. In addition, individualizing foundations show weaker but still directionally supportive associations: Fairness is positively associated with the amplifier role (b = 0.15) and negatively with the attenuator (b = −0.07). Care shows a small positive coefficient for the amplificatory role (b = 0.015) and a negative one for the attenuator role (b = −0.019). These patterns suggest that users who emphasize individualized moral expressions—fairness and care—are more likely to be amplifiers, spreading toxicity further. Although not all effects are large. In summary, people with binding moral foundations overall tend to manifest a high likelihood of being amplifiers, except for loyalty, which actually increases the odds of being attenuators. Contrary to prediction, people with high individualized moral expressions (i.e. care and fairness) are more likely to be amplifiers and less likely to be attenuators. Notably, users with fewer followers and retweets remain more likely to act as amplifiers, suggesting that amplification of toxicity may be driven by less prominent accounts (see Supplementary Appendix F).
The H3 posited that users’ network positions would shape their roles in the diffusion of toxicity. Specifically, we expected that highly embedded users would be more likely to attenuate toxicity; users positioned as structural bridges would function as amplifiers; and users embedded in close-knit communities would tend to be copycats. This model (Model 3) achieves less accurate results than Model 2 (precision = .76, accuracy = .59). The findings partially support H3, with notable deviations. Contrary to expectations, amplifiers were not associated with structural bridges. Instead, amplification of toxicity was more strongly predicted by high eigenvector centrality (b = 0.13), PageRank (b = 0.337), and out-degree (0.369)—all indicators of network embeddedness. This suggests that users who are prominent and active within the network, rather than peripheral connectors, are more likely to amplify toxicity. These results point to a more centralized diffusion dynamic, where core users—not bridges—serve as the primary amplifiers.
The profile of attenuators partially aligns with H3. Attenuators were associated with high in-degree (b = 11.69) and eigenvector centrality (b = 0.16), consistent with the notion that embedded users may play a moderating role. However, the positive effect of betweenness (b = 0.25) and the negative association with PageRank (b = −0.22) complicate this picture. One possible interpretation is that attenuators may occupy selective forms of embeddedness—well connected in terms of incoming ties, but not necessarily among the most visible or influential accounts in terms of reach.
Copycats showed higher ego density (b = 0.18) and reciprocity (b = 0.03), supporting the idea that imitation tends to occur in dense, mutually interactive ego networks. However, the weaker associations with transitivity (b = −0.007) and closeness centrality (b = −0.05) suggest that these users may not belong to fully cohesive communities. Instead, their mimicry may emerge within moderately cohesive clusters where interaction is frequent but not closed.
The RQ explores which type of predictors most strongly shape users’ likelihood to be amplifiers, attenuators, and copycats. We first estimated a model including all variables (precision = .82, accuracy = .64). Overall, for amplifiers (Model 4), we found that they are more likely to emphasize fairness and sanctity/moral purity. They are typically highly embedded and influential actors in the network, as indicated by higher PageRank, eigenvector centrality, and out-degree. In contrast, users who receive many incoming ties are significantly less likely to be amplifiers. Similarly, users with a more pro-immigration stance are less likely to amplify toxicity. High levels of retweet activity may signal a more passive role, also associated with a lower likelihood of amplification. Interestingly, users with more followers are less likely to amplify toxicity, suggesting that prominent users may avoid amplifying toxic content.
Attenuators, on the other hand, are more likely to hold pro-immigration stances and to receive a high number of ties. Network position also plays a role—betweenness centrality is positively associated with attenuation, indicating that structural bridges may play a role in mitigating toxicity. Having more followers also increases the likelihood of attenuation. Conversely, users who emphasize sanctity or fairness are less likely to be attenuators. In addition, users with lower out-degree and PageRank—less active or visible users—are more likely to attenuate. A denser ego network decreases the likelihood of attenuation.
Copycats are more likely to be embedded in dense, reciprocated networks, as indicated by high ego density and reciprocity. They also receive many incoming ties and are likely to mimic others’ behaviors through retweeting. Users with a moderate number of followers are more likely to fall into this group. In contrast, copycats are not driven by sanctity or authority moral concerns and are less likely to occupy central (low PageRank) or bridging (low betweenness centrality) positions.
In addition, we also rerun using random forest regression (precision = .85, accuracy = .80). Figure 1 illustrates the feature importance of the random forest model. The results reveal that psychological and network-related features dominate predictive power. Specifically, sanctity emerged as the most influential predictor, followed closely by page rank, closeness, and in-degree, which capture centrality and connectivity within the network. This suggests that both moral language and structural prominence are key indicators driving user toxicity diffusion. In contrast, traditional engagement metrics such as retweet count, likes, and quotes showed comparatively lower importance, as did the immigration stance variable.
Full model comparison of feature importance associated with different predictors.
Discussion
Toward an empirical typology of antecedents of toxicity diffusion
Informed by findings, we present a typology to explain the antecedents of toxicity diffusion on social media by integrating psychological dispositions, network structures, and behavioral engagement histories. We integrate moral foundation theory and the network diffusion literature to explain patterns of toxicity diffusion in online immigration discourse. Our typology assumes role-based diffusion mechanisms and delineates three distinct roles in the propagation of toxicity. Amplifiers are highly central users, often driven by moral convictions, who actively spread uncivil content. Their structural embeddedness and moral motivations render them potent disseminators. Attenuators, in contrast, resist or discourage toxicity, a tendency shaped by fairness-oriented morality or by peripheral network positions that reduce exposure and pressure to engage. Copycats replicate toxicity after exposure, typically triggered by structural opportunities for imitation.
This empirically based typology contributes to theorizing in several ways. First, this typology shows connections between the morality foundation theory and network theories, reflecting that the three domains jointly shape toxicity diffusion. The first domain comprises individual-level psychological antecedents. Our results support theoretical expectations from MFT (Haidt and Graham, 2007) and extend prior research on the link between moral language and online civility (Kim et al., 2022). Findings also reveal that different moral dimensions are related to toxicity network roles, pointing to a future theorizing direction of integrating the psychological-structural interdependence.
Specifically, binding foundations, as discussed in the literature, often function to uphold group cohesion by condemning norm violations (Baldner and Pierro, 2019). In immigration debates, sanctity concerns may be especially prominent, with high scorers perceiving immigration as a threat to cultural purity or national integrity, justifying toxicity as morally acceptable. It also reflects fears that demographic, cultural, and linguistic changes are challenging family and community values. These fears may generate defensive, sometimes hostile, rhetoric. The positive link between sanctity and amplifier behavior suggests this moral foundation may serve as a cover for inflammatory rhetoric that defends in-group values and penalizes out-group behavior. The weak effect of authority suggests that authority is less central than other moral foundations in driving toxicity diffusion in this context. Although the effects are weak, authority does show positive associations with being both amplifiers and attenuators, rather than passively mimicking their neighbors’ toxicity levels. This may reflect that some individuals with authority amplify toxicity toward norm violators, such as justifying enforcement and boundary policing, while others uphold norms of civil discourse and institutional legitimacy, thereby attenuating toxicity.
The second domain concerns network positioning, where findings both confirm and challenge existing network diffusion theory. A user’s location in the social network influences both exposure to toxicity and the capacity to diffuse it. Contrary to expectations that bridging positions would facilitate toxicity diffusion (Hopp and Vargo, 2019), our findings suggest that amplification is driven by deeply embedded users (reflected in low betweenness centrality and low closeness centrality, which means they were not likely to be brokers, and hard to reach others in the network). This points to a centralization dynamic in which influential users connected to well-connected others (reflected in high out-degree centrality and page rank) within echo chambers serve as key spreaders of toxicity. These users can shape discourse agendas and norms, making their incivility more visible and more likely to be emulated. This challenges the assumption that bridge actors are the primary vectors of toxic discourse, highlighting instead how visibility (Schulze et al., 2024), motivated outreach, and access to prestigious others in networks condition toxicity amplification.
Copycats are active and popular participants, as reflected in their high centrality. However, their overall network impact is low. They are also not brokers in communities, despite their mimicry and easy-to-influence community behaviors. They are also occupying network positions that are not easy to reach others and are not closely connected with network elites. They seem to function as local community glue, reinforcing dense conversations, but their influence is limited in motivating their neighbors to form conversations with one another or connecting disconnected global structures. Attenuators occupy central bridging positions and have high access to others in the network, suggesting their efforts to connect disconnected echo chambers and reduce toxicity spread. They are often popular, but restrained from spamming send behaviors. Their high bridging efforts also lead to low efforts in creating local norms and clustered communities, which may limit the long-lasting effect of their intervention efforts. They are also not especially connected with prestigious and well-connected actors, either, which may further limit their impact in the network.
The third domain, user engagement patterns, although not the focus of hypotheses, also contributed meaningful insights. We found that visibility may encourage caution or accountability, while users with smaller audiences might feel freer to be provocative or conform to toxic norms, echoing Schulze et al. (2024)’s finding that anonymity and visibility are critical platform affordance factors when it comes to social media toxicity. In addition, the negative link between retweet count and amplifier status indicates that highly toxic content may not always go viral in this context. Together, these three domains—psychological dispositions, network structures, and engagement histories—offer a comprehensive account of how and why toxicity spreads in online political discourse. None of the factors alone fully explains the toxicity diffusion roles. As further supported by the random forest analysis, the most important factors include variables from all three domains—sanctity (psychological dispositions), page rank, closeness, in-degree centrality (network structures), and followers count (engagement profiles). A complete theoretical account, therefore, should include all three dimensions when analyzing the diffusion of toxicity. By theorizing the roles of amplifiers, attenuators, and copycats, this typology advances our understanding of digital toxicity not as an undifferentiated contagion but as a structured, value-driven, and network-mediated phenomenon.
Moreover, the typology reveals that empirical findings contradict predictions from established theories, pointing toward the need for theoretical refinement. To start, interestingly, loyalty, although a binding foundation, correlates more with a high likelihood of being an attenuator. Although people who emphasize loyalty in their expressions also tend to emphasize group boundaries, they remain very protective of their group norms. In the immigration context, these people defend group identities and shared values. Toxicity would not be strategic or consistent with legitimizing their loyalty arguments. Existing studies connecting loyalty and the style of communication are almost nonexistent, and future online discourse research should fill the gap and explore more outcomes for emphasized loyalty expression, especially differentiating the analysis of their in-group and out-group communication (Kim and Park, 2019).
In addition, contrary to prediction and expectation, individualizing moral foundations (i.e. care and fairness) play an amplifying rather than an attenuating role. This may reflect that conversations emphasizing care and fairness, although directly linked with discussions of individual rights and well-being, might be a polarized, highly emotional, and contested space. We conducted further post hoc comparisons to understand the different moral influences between anti- and pro-immigration stances and found potential toxicity triggers (see Supplementary Appendix F). These contradictions reflect that the connection between morality and toxicity spread is not as simple as binding and individualizing mapping. The MFT specifies the relationship between foundations and attitudes, but inadequately addresses the connection between foundations and communication behaviors in contested discourse networks. Future research should delve deeper into the moral strategies and the spread of toxicity to specify the theoretical mechanisms. For example, future research could explore whether moral foundations, situated in contested and emotionally charged conversations, may lead to intensified, rather than reduced, conflict and toxicity. The interesting counterintuitive finding about loyalty serving as norm enforcement rather than group-boundary policing (i.e. further emphasizing group boundaries and excluding out-groups) shows the possibility for future research to dissect different loyalty expression strategies and understand mechanisms leading to norm enforcement versus further group separation.
In all, this article reflects that moral foundations shape not only attitudes but also behavior in contested digital environments. As Hwang et al. (2018) argued, exposure to toxicity can trigger emotionally charged responses such as moral outrage and closed-mindedness. When processed through the lens of most binding foundations, these emotions may escalate into aggressive, group-protective speech. Conversely, individualizing foundations may foster empathic engagement and moral inclusion, thereby dampening reactive toxicity.
Notably, copycats have relatively weaker moral profiles in comparison to amplifiers and attenuators. They do have strong local network embeddedness (high ego density and high reciprocity), but weak global network integration and impact (low in transitivity, not bridges, and low overall centrality except for in-degree centrality). These may reflect that local embeddedness in cohesive communities may suppress moral expressions, which may further limit global network impact.
The influence of moral foundations is further conditioned by issue stance. A stronger pro-immigration position was associated with lower odds of becoming an amplifier and higher odds of becoming an attenuator, relative to copycats. This pattern suggests that communicative behaviors in toxic online environments are not merely reactive but shaped by underlying ideological values. As expected, anti-immigration users were more likely to escalate toxicity, while pro-immigration users tended to exercise greater restraint, reflecting their emphasis on empathy and humanitarian concerns (Stravato, 2024). Neutral users, lacking clear ideological commitments, appeared more susceptible to mimicking the prevailing tone of their environment, consistent with the copycat role.
In summary, this empirically based typology serves as an intermediate contribution between descriptive empiricism and a theoretical framework. It advances beyond mere descriptive and empirical contributions by (1) identifying connecting pathways between moral foundation theories to network diffusion research through toxicity diffusion roles and (2) identifying consistencies and inconsistencies in current theorizing and proposing potential directions toward more solid theoretical framework development.
Profiles of amplifiers, copycats, and attenuators
Our analysis identifies distinct structural and behavioral configurations associated with the roles of amplifiers, copycats, and attenuators in the diffusion of toxicity. Building on these empirical patterns, we develop an empirically grounded typology that characterizes each role across three domains: psychological dispositions, network structures, and engagement profiles (see Table 2). This typology synthesizes observed regularities rather than positing a fully specified causal model, providing a structured foundation for future mechanism-driven theorizing.
Empirically grounded typology of three toxicity diffusion roles.
Features included in this table are those statistically significant predictors from Model 4, as well as those of high importance in the random forest analysis.
Among them, amplifiers exhibit a structural profile that challenges prevailing assumptions about the peripherality of toxic actors. Rather than operating from the margins, these users are centrally embedded within the network. Their position enables them to serve as hubs of content dissemination, with structural influence that allows toxic messages to be widely circulated. The prominence of moral concerns around sanctity among these users suggests that amplification is not only a function of network embeddedness but may also be motivated by perceived violations of deeply held values—particularly those associated with moral outrage. At the same time, the negative relationship between amplification and follower count indicates a balancing act: users with broader audiences may temper their amplification of toxicity to manage reputational risks. This combination of moral motivation and strategic calculation underscores the dual forces that drive the spread of toxic content from the network’s core.
In contrast, attenuators reflect a more complex and selective form of structural centrality. Consistent with our theoretical expectations, attenuators tend to have high in-degree and eigenvector centrality, suggesting that their visibility within the network—by virtue of receiving attention from many others—may create incentives to conform to prevailing norms and avoid overt incivility toxicity. However, the positive relationship between betweenness centrality and the negative association with PageRank indicate that attenuators are not uniformly influential across the network. Instead, they appear to occupy key bridging positions, embedded within clusters where civility may be more strongly reinforced through local norms. These users are likely exposed to diverse perspectives and may serve as buffers against the unfiltered spread of incivility, exercising restraint not only out of personal disposition but also due to the discursive expectations of their network surroundings. The observed alignment between attenuation, a pro-immigration stance, and high follower counts further suggests a convergence of normative commitments and reputational considerations in discouraging the amplification of toxicity. It is also noteworthy that Schmid et al. (2024) found that users may choose not to engage with toxic content because they are consciously aware that such engagement would increase the visibility of harmful content. This suggests that being an attenuator may reflect strategic and active decision-making, rather than merely the passive occupation of particular network positions. This raises an important question that our data could not directly address and that warrants future research: do attenuators consciously and deliberately restrain engagement as an informal moderation strategy, or does their central network position heighten their awareness of toxicity diffusion and thereby motivate more reputable behavior?
Copycats, by comparison, operate at a different level of the network hierarchy and exhibit behavioral patterns driven more by social mimicry than by moral considerations. Their high ego network density and reciprocity suggest they are embedded in tightly-knit relations, where conformity to group norms—including toxicity—can be amplified through repeated mutual interactions. These users are not necessarily located in cohesive or highly central subgraphs; the weak associations with transitivity and closeness centrality imply that full community integration is not a prerequisite for mimicry. Instead, frequent and reciprocated interactions appear sufficient to promote the imitation of tone and behavior. Copycats tend to be reactive, with high retweet activity and low scores on PageRank and betweenness centrality, indicating their limited influence in steering discourse but high susceptibility to it. Their behavior underscores the performative and relational dimensions of toxicity, where tone becomes a function of peer behavior within localized contexts rather than broader ideological commitments.
Theoretical and practical implications
This study carries significant implications for the theoretical understanding of online toxicity, a subtype of incivility that focuses on explicit individual-level incivility, its networked diffusion, and the relational dynamics that sustain or inhibit it. First, by identifying distinct roles—amplifiers, copycats, and attenuators—and linking them to structural positions within the network, this research moves beyond individualized psychological explanations and toward a relational and positional account of behavior. Existing theories of toxicity often assume a dichotomy between ideologically motivated actors and disengaged participants; our results suggest instead a multidimensional ecology in which users play differentiated roles shaped as much by their location within the network as by their normative commitments or dispositions.
Moreover, the findings refine diffusion models by illustrating that not all diffusion is driven by influence in the traditional sense; instead, mimicry and local conformity, particularly in dense ego networks, can sustain toxicity without requiring ideological alignment or centralized coordination. These insights challenge linear contagion models and call for greater attention to the performative and socially contingent nature of communication behaviors online.
The study highlights the limitations of interventions that target toxicity solely at the level of individual behavior or platform content moderation. Efforts to reduce toxic discourse must also consider the relational and structural dynamics that enable its spread. Platform designers and policymakers should be attentive to the role of highly central actors in amplifying toxicity and consider strategies that mitigate the structural rewards associated with toxic amplification—such as algorithmic adjustments that reduce the visibility of uncivil content from high-centrality users. At the same time, supporting attenuators—users who function as informal gatekeepers—may offer a bottom-up strategy for reinforcing discursive norms. Encouraging bridge actors through reputation-building tools or by promoting diverse exposure may help strengthen containment dynamics within platforms. Finally, the role of copycats suggests the importance of disrupting microlevel social reinforcement loops. Interventions at this level could involve fostering counter-norms in tight-knit groups, using nudges or visual cues to signal civility as a normative standard. Educational initiatives that focus on digital literacy and relational awareness—teaching users how their engagement patterns shape the tone of discourse—could also help counteract the mimicry of toxicity. This research arrives at a critical time of platform governance debates, marking X’s shifts from human-powered content moderation to reduced moderation to algorithm-mediated content moderation under Musk’s leadership. The tension between Musk’s advocacy for free speech and X’s struggle to cope with harmful content points to social media platforms’ insufficient moderation strategies and efforts. When harmful content is pushed to the margins, they do not disappear. Instead, they appear in a more covert way, creating challenges for people to recognize their harms (Rieger et al., 2021). Our research suggests that more analytical efforts should be focused on network and psychological positions based on user content, providing a more nuanced lens for understanding toxicity diffusion mechanisms.
Limitation and future research
We acknowledge that the data only captures a small period of immigration policy shifts. We also encourage future research to examine how stable these roles are over time and under what conditions users transition from one role to another. Longitudinal studies that track changes in users’ network positions, exposure to toxicity, and evolving behavioral patterns could deepen our understanding of how these dynamics unfold. Detoxify does not capture the targets of the toxicity. For example, toxicity that attacks individuals and targets specific racial groups would potentially differ in patterns of spread as they may instigate different levels/types of emotions. We encourage future research to dive deeper into the nuanced toxicity targeting patterns. Another fruitful avenue involves integrating psychological measures—such as moral foundations—with network metrics to develop more nuanced predictive models. While this study provides indirect evidence of the interaction between moral concern and network position, future research could benefit from experimentally or observationally isolating the effects of moral framing in relation to users’ structural affordances. We also acknowledge limitations in data coverage. Our data begins in 2021, when Biden assumed office; however, immigration is a long-standing social issue whose salience and complexity become especially pronounced during presidential election cycles. Our team plans to continue expanding data collection to include Trump’s first and second administrations in order to capture a broader range of events and policy debates. There is also an inherent limitation associated with retrospective data scraping, which may lead to missing data in nonrandom ways. For example, more harmful, extreme, or biased posts may be more likely to be removed by the platform or moderators, potentially producing a distorted representation of the distribution of toxicity in our data. Future research should compare real-time data collection with retrospective scraping and examine how content removal interacts with toxicity levels, as this process may introduce systematic bias into the analysis.
Generalizability
Due to data access constraints, we were unable to collect retweet or follower–followee networks for comparative analysis. As a result, our findings are specific to mention networks and therefore capture proactive, target-directed expressions of toxicity. Mentions reflect actor-initiated communication in which users deliberately engage others, often in ways that signal moral positioning or interpersonal confrontation.
In contrast, retweets represent a distinct diffusion mechanism. Reposting may involve passive amplification of toxic content without generating original expression or direct engagement. These interaction modalities likely rely on different psychological processes. Toxicity amplification through mentions may be driven by intense moral conviction and emotional arousal—such as outrage—whereas passive amplification via retweets may not require the same degree of affective activation. Consequently, our findings should not be generalized to message-driven diffusion processes such as repost cascades.
Future research should systematically compare how different network modalities—mentions, replies, retweets, and follower structures—shape amplification and attenuation dynamics. In particular, scholars could examine whether moral foundations predict not only proactive toxicity expression but also selective reposting behaviors (e.g. Kumar and Xiao, 2022). In addition, the role of repost cascades in sustaining or escalating toxic discourse warrants further investigation. Understanding how mentions and reposts jointly contribute to norm formation, reinforcement, and potential escalation may provide deeper insight into pathways for interrupting toxic diffusion in online communities.
Finally, users might exhibit different diffusion roles in different social networks. For example, a user could be a copycat in retweet networks, but showcases strong amplification behaviors in mention networks. The current classification does not consider the whole toxicity spread roles across social media networks, and future research should explore whether patterns persist with a more comprehensive mapping of people’s toxicity behaviors. This focus also limits practical implications in platform governance. Our findings only apply to understanding interventions targeting proactive toxic communication. Future efforts should be spent on identifying more comprehensive toxicity reduction strategies through content moderation on all types of content, algorithmic adjustment, and reward systems for attenuation.
Beyond network type, this study is situated within a specific platform and historical context: X (formerly Twitter) during the Biden administration, following Musk’s acquisition and substantial reductions in trust and safety staffing. This period was characterized by comparatively lower levels of centralized content moderation. Public disclosures under the E.U. Digital Services Act indicate that X’s moderation staff-to-user ratio (1:60,249) is substantially lower than that of other major platforms such as TikTok, LinkedIn, and Meta (Hutchinson, 2024). Earlier reporting noted that Facebook alone employed more than 15,000 content moderators (Beijbom, 2022). Concurrently, reports suggest that hate speech increased markedly following the ownership transition (Brewster, 2024; Hickey et al., 2025).
These governance conditions likely shape the opportunity structure within which amplification, attenuation, and mimicry unfold. In an environment with reduced moderation and enforcement, users may face fewer institutional constraints on toxic expression, potentially increasing amplification dynamics while weakening deterrents that might otherwise support attenuation. However, because we do not conduct cross-platform comparisons, we cannot causally attribute observed diffusion patterns to moderation intensity. Our claims should therefore be interpreted as conditional on a relatively low-moderation environment.
Platform affordances further define the boundary conditions of our findings. The X is characterized by (1) short-form content (280 characters for most users during the study period), (2) public-by-default visibility, (3) direct-mention notifications to targeted users, and (4) algorithmic ranking systems that privilege engagement-generating content. Short-form communication may reduce contextual nuance and accelerate reactive exchanges, particularly on contested political issues. Public visibility increases reputational stakes and cascade potential. Mention notifications facilitate direct interpersonal escalation. Engagement-driven algorithms may disproportionately surface emotionally charged and morally expressive content, which prior research shows is more likely to diffuse broadly (Brady et al., 2017).
These affordances do not mechanically cause toxicity diffusion patterns, but they create structural conditions under which morally charged amplification, especially those rooted in binding foundations such as sanctity, may scale rapidly. Accordingly, our identified profiles of amplifiers, copycats, and attenuators should be understood as emerging within a platform architecture that combines high visibility, rapid interaction cycles, engagement-based ranking, and comparatively limited moderation.
Generalization to other platforms should therefore be made cautiously. Platforms with longer-form content, stronger group segmentation, heavier moderation, or different algorithmic logics (e.g. Facebook/Meta, Reddit, and TikTok) may alter the relative prevalence or behavioral expression of these roles. For example, toxicity may be amplified within closed or semiclosed communities, yet encounter barriers to platform-wide diffusion under stricter moderation regimes. Cross-platform comparative research is necessary to test whether the amplifier, copycat, and attenuator profiles reflect generalizable social-psychological dynamics or are partly contingent on specific platform affordances and governance structures.
Finally, this research is situated within the immigration issue domain, which raises important questions about issue-specificity and generalizability. Prior moral foundations research suggests that different political issues foreground different moral concerns. For example, purity—associated with spiritual contamination and adherence to a perceived natural or social order—has been shown to be especially salient in debates about abortion (Lockhart et al., 2023). Climate change discourse frequently invokes care and fairness concerns, and in some cases purity-related themes (Dickinson et al., 2016). Gun control debates often center on care (harm prevention), fairness (rights and justice), and liberty. Like immigration, these issues are highly partisan and characterized by emotionally charged, confrontational exchanges (Dickinson et al., 2016; Hansen and Seppälä, 2024; Lockhart et al., 2023).
Given this variation in moral salience across issues, we do not assume that the specific moral foundations predicting toxicity amplification in immigration discourse would necessarily operate in identical ways in other domains. Rather, we propose that the broader mechanism linking moralized language to toxicity amplification—particularly when moral claims are embedded in polarized network structures and reinforced through engagement dynamics—may generalize across contentious topics. However, the specific moral foundations most strongly associated with toxicity spread are likely to vary by issue depending on which moral concerns are culturally and politically salient in that domain. We encourage future research to adopt a comparative, cross-issue design to test whether the observed patterns replicate across debates such as abortion, gun control, and climate change, and to identify issue-contingent versus domain-general processes in toxic content diffusion.
Supplemental Material
sj-docx-1-nms-10.1177_14614448261447866 – Supplemental material for Who fuels the fire? Predicting amplifiers, copycats, and attenuators in the networked spread of toxicity in immigration discourse
Supplemental material, sj-docx-1-nms-10.1177_14614448261447866 for Who fuels the fire? Predicting amplifiers, copycats, and attenuators in the networked spread of toxicity in immigration discourse by Yiqi Li and Aimei Yang in New Media & Society
Footnotes
Funding
The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This paper is funded by Syracuse University, School of Information Studies.
Declaration of conflicting interests
The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Supplemental material
Supplemental material for this article is available online.
Author biographies
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
