Affective polarization and dynamics of information spread in online networks

In the United States, Democrats and Republicans increasingly not only disagree on policy but also express mutual dislike and distrust, a phenomenon termed affective polarization. This emotional divide has been implicated in diminishing cross-party cooperation, increasing hostility toward out-groups, and eroding trust in institutions, as exemplified during the COVID-19 pandemic when partisan identity shaped responses to public health guidance. While multiple drivers of affective polarization have been proposed, the role of social media remains contested. Social platforms can amplify inflammatory language and moral outrage and may foster echo chambers that concentrate like-minded users, potentially exacerbating polarization; at the same time, exposure to out-party views persists and can, paradoxically, worsen polarization. Against this backdrop, the study asks how to quantify affective polarization in social media interactions and whether it is tied to the structure of online networks. The authors propose an instrument that measures affect via emotions and toxicity expressed in reply interactions between users with known ideology. They hypothesize that in-group replies will be more positive, while out-group replies will be more negative and toxic, and that emotional expression will vary systematically with network distance between interacting users. Further, they investigate whether emotionally polarized groups exhibit distinct patterns in the diffusion of information about contentious issues. The paper’s contributions are three-fold: (1) introducing a measurement approach that focuses on how people talk to, rather than about, out-group members; (2) establishing an empirical relationship between emotions and network distance, treating affective polarization as a structural property of online networks; and (3) characterizing partisan differences in the persistence and dynamics of information spread on Twitter surrounding the COVID-19 pandemic and the abortion debate.

Prior work attributes affective polarization to influences from news media, political elites, and demographic factors. The role of social media is debated, but several studies suggest it amplifies moralized and inflammatory content, particularly out-group animosity, and can create echo chambers that increase polarization. However, users are not fully isolated—cross-cutting exposure occurs and may intensify polarization. Existing measures of affective polarization often examine how individuals speak about out-groups; this study instead focuses on direct interpersonal interactions (replies). The work builds on research demonstrating the emotional drivers of engagement and the networked nature of online discourse, leveraging advances in transformer-based language models for emotion and toxicity detection, and prior approaches to estimating user ideology from shared URLs and text. It also connects to scholarship on information diffusion, community structure, and the differential spread of content in polarized environments.

Data: The study analyzes two public Twitter corpora. (1) COVID-19: tweets posted between January 21, 2020 and April 22, 2020; (2) Abortion (Roe v. Wade): tweets from January 1, 2022 to January 6, 2023 spanning the period before and after the Dobbs decision. Tweets were filtered to users located in the United States using the Carmen geolocation tool, which infers location from tweet metadata and user bios. Interactions considered were retweets and replies where both the author and the parent tweet’s author were located in the U.S. Dataset statistics: Roe_v_Wade—7,131,980 retweets (1,005,156 retweeters) and 460,868 replies (172,988 responders); COVID-19—46,419,871 retweets (10,758,690 retweeters) and 4,173,679 replies (833,875 responders). Ideology classification: Users were classified as liberal or conservative using a text-embedding-based classifier trained on labels derived from Media Bias-Fact Check (MBFC) political bias scores of URLs users shared. Each user received a weighted average score over shared sources to create training data, achieving state-of-the-art performance per prior work. Emotion and toxicity detection: Emotions in replies were detected with an open-source transformer-based multi-label model (Demux/SpanEmo), which outputs confidence scores for emotions including anger, disgust, fear, sadness, joy, love, trust, pessimism, and optimism. Toxicity was measured with Detoxify, which outputs a scalar likelihood of toxic content; only the toxicity score was used. Measuring affect in interactions: In-group interactions are replies between same-ideology users; out-group interactions are replies between opposite-ideology users. The study computed the distribution and means of emotion confidence scores and toxicity for these interaction types, comparing groups using two-sided Mann–Whitney U tests with Bonferroni correction. Reply length (characters normalized by 280) and word count were also analyzed. Network construction and distance: Retweet-based social networks were built for each dataset as undirected, unweighted graphs: nodes are users; an edge indicates at least one retweet in either direction, signaling exposure to the same information. Networks were embedded into a lower-dimensional space using LargeVis for visualization and to compute Euclidean distances between users. Network distance between interacting users was measured in two ways: (1) Euclidean distance in the embedding space (normalized to [0,1]); (2) shortest path length in the retweet network. Distances were binned into quartiles (Q1 closest to Q4 farthest) for analysis. Regression analysis: Linear regressions modeled emotion/toxicity confidence as a function of normalized network distance between interacting users. Regressions were conducted overall, separately for in-group and out-group interactions, and after disaggregating by ideology. Robustness checks compared results using embedding-based distance versus shortest-path distance. Issue detection and diffusion dynamics: Contentious issues were detected using transformer-based classifiers trained on labels derived from Wikipedia-harvested keywords/phrases. COVID-19 issues: origins, lockdowns/social distancing, masking, education/online learning, vaccines. Abortion issues: religion, fetal rights, exceptions to bans, women’s health, bodily autonomy. For each ideological group, daily retweet volumes per issue were standardized via z-scores and analyzed over 80-day windows. Temporal dynamics were characterized using the autocorrelation function (ACF); the longest significant lag (time at which ACF fell below confidence bounds) served as a measure of persistence. Comparisons were made across time periods (e.g., early pandemic vs. summer 2020; pre-leak vs. post-Dobbs decision).

Affective polarization in replies: Across both datasets, out-group replies exhibited higher anger, disgust, and toxicity than in-group replies, while in-group replies showed higher joy; these differences were statistically significant. Fear behaved differently from other negative emotions, tending to be higher in in-group replies, consistent with a social cohesion role of fear. Out-group replies were shorter: in Roe_v_Wade, out-group replies were on average 17 characters shorter; in COVID-19, 8 characters shorter, despite being more emotional, suggesting cross-party replies often express animosity or trolling rather than information sharing. Structural property with network distance: Emotions varied systematically with retweet-network distance. As distance increased (Q1 to Q4), anger, disgust, and toxicity increased, while joy and fear decreased; reply length also decreased with distance. Linear regression coefficients confirmed these trends across both datasets. Results held separately within in-group and out-group interactions and after disaggregating by ideology. Findings were robust to the choice of distance metric (embedding distance vs. shortest-path length). Network structure: The Roe_v_Wade retweet network showed two large polarized communities; the COVID-19 network (early pandemic) exhibited a multi-focal structure, with less entrenched polarization at that time. Information diffusion dynamics by group and issue: In the early pandemic, liberals showed persistent attention to education/online learning, while conservatives showed persistence for lockdowns. In summer 2020, masking became a salient and persistent issue for both groups. In abortion discussions, after the Dobbs ruling, conservatives exhibited increased persistence around religion, fetal rights, and exceptions to abortion bans, compared to the pre-leak period. These dynamics indicate that emotionally salient issues to a group exhibit longer-lived diffusion patterns. Scale: Analyses covered tens of millions of retweets (COVID-19: 46.4M; Roe_v_Wade: 7.1M) and millions to hundreds of thousands of replies (COVID-19: ~4.17M; Roe_v_Wade: ~461K), strengthening confidence in observed patterns.

The results validate that direct interpersonal interactions on social media manifest affective polarization: people communicate more positively within ideological in-groups and more negatively and toxically across ideological lines. Crucially, by tying emotional expression to network distance, the study shows that affective polarization is not solely a function of explicit group labels but also a structural property of online networks: users express warmer emotions toward nearby network neighbors and colder, more hostile emotions toward distant ones. This implies that even absent partisan labels, polarization can be inferred from network topology and interaction patterns. The distinct persistence of issue diffusion within ideological groups suggests that emotional salience modulates attention and the longevity of information flows, shaping agenda-setting and public discourse. The observed differences between replies and retweets underscore that these interaction types serve different communicative roles; conflating them may obscure key features of network structure and affect. Together, these findings illuminate how emotions, network structure, and partisanship interact to drive polarized engagement and information spread.

This study introduces an instrument to quantify affective polarization from reply interactions, demonstrates that emotional valence correlates with network distance—generalizing beyond the in-group/out-group dichotomy—and shows that emotionally salient issues diffuse with different temporal dynamics across ideological groups. The findings are consistent across large-scale datasets on COVID-19 and abortion, highlighting both the measurement value of replies and the structural nature of affect in online networks. Future research should (1) replicate analyses on broader, more representative and cross-platform datasets; (2) disentangle causal directionality between network distance and emotive expression; (3) better control for potential confounders, including algorithmic personalization and content selection effects; (4) incorporate richer network representations (e.g., follower graphs, weighted/multiplex ties) and longitudinal network evolution; and (5) examine interventions or design changes that might mitigate toxic cross-group engagement while preserving exposure to diverse viewpoints.

The datasets, while large, represent only portions of online discussions and are subject to sampling and keyword-based collection biases. Not all interactions are observable, potentially limiting the range of detected emotions. Retweets approximate, but do not equal, follower relationships, which may affect network-based conclusions. Classifiers for ideology, emotions, and toxicity can introduce errors. The study is descriptive and does not identify causal directionality between network distance and emotional expression. Confounding factors—such as emotionally charged content being more likely to be retweeted, or platform personalization amplifying emotional content—may also contribute to observed patterns. Despite these constraints, consistency across datasets, measures, and robustness checks lends confidence to the conclusions.