Toxic behavior silences online political conversations

TL;DR

This paper addresses how toxic online behavior may silence political conversations by inducing self-censorship. It combines a dataset of about $32.5$ million comments from six US YouTube outlets during Sep 2020–Apr 2021, measures toxicity and insults with Perspective API, and analyzes sequential dynamics with a two-state Hidden Markov Model characterized by latent states $Z_1$ and $Z_2$ and emission probabilities $P(X_j|Z_i)$. The results show that toxicity and insults spike around major political events and that a terminal state $Z_1$ is associated with higher toxicity and end-of-conversation dynamics, while $Z_2$ corresponds to earlier engagement with lower toxicity. Topic-based clustering via hierarchical SBM confirms robustness across groupings, revealing topic-dependent toxicity/insult patterns, such as heightened insults for social-justice topics near termination and elevated toxicity for COVID-19 topics in certain clusters. These findings underscore self-censorship as a mechanism shaping political discourse online and suggest platform interventions to foster diverse, constructive deliberation.

Abstract

Quantifying how individuals react to social influence is crucial for tackling collective political behavior online. While many studies of opinion in public forums focus on social feedback, they often overlook the potential for human interactions to result in self-censorship. Here, we investigate political deliberation in online spaces by exploring the hypothesis that individuals may refrain from expressing minority opinions publicly due to being exposed to toxic behavior. Analyzing conversations under YouTube videos from six prominent US news outlets around the 2020 US presidential elections, we observe patterns of self-censorship signaling the influence of peer toxicity on users' behavior. Using hidden Markov models, we identify a latent state consistent with toxicity-driven silence. Such state is characterized by reduced user activity and a higher likelihood of posting toxic content, indicating an environment where extreme and antisocial behaviors thrive. Our findings offer insights into the intricacies of online political deliberation and emphasize the importance of considering self-censorship dynamics to properly characterize ideological polarization in digital spheres.

Figures (9)

• Figure 1: Trends in comment sentiment and volume over time. The figure shows daily sentiment proportions (blue line), total number of comments per day (gray line), 7-day rolling average (red line) and relevant events annotated with black dots. (a) Proportion of toxic comments, defined as those with a toxicity score greater than 0.6. Peaks in toxicity are observed around September 2020, coinciding with heightened social and political tensions following the Black Lives Matter (BLM) protests. A subsequent peak occurs in January 2021, likely linked to the increased tensions surrounding the Capitol Attack. (b) Proportion of insulting comments, defined as those with an insult score greater than 0.6. The trend in insulting comments resembles that of toxicity; however, there is a more pronounced peak in February 2021, likely related to events surrounding the COVID-19 pandemic. (c) Total number of comments per day. A notable surge in activity coincides with the controversial events of the 2020 election; but, this increase is not mirrored in the trends for toxicity or insulting comments. The highest comment counts during the January 2021 Capitol Attack coincide with spikes in both toxicity and insults.
• Figure 2: Probability density functions of sentiment scores for replies ($Y$) based on the sentiment scores of top-level comments ($X$). The dotted line represents the unconditional distribution $P(Y)$, showing the overall probability density of the sentiment scores for replies, while the solid lines indicate the conditional probabilities $P(Y|X)$ for various ranges of $X$. Both panels show a correlation between the negative sentiment of top-level comments and their responses, with replies to highly toxic/insulting comments (brown lines) being more likely to have high toxicity/insult scores (greater than 0.8) than other replies (a) Probability density functions of toxicity scores. Replies are more likely to have toxicity scores between 0.2 and 0.6 than to have insult scores in the same range. (b) Probability density functions of insult scores. Replies are more likely to have low insult scores ($<$0.2) compared to low toxicity scores.
• Figure 3: Diagram of the fitted Hidden Markov Model. The model has two latent states (${Z}_{1}$ and ${Z}_{2}$) and three observed signals (${X}_{1}$, ${X}_{2}$ and ${X}_{3}$). The arrows represent the transition probabilities between latent states ($P({Z}_{t,i}|{Z}_{t-1,i})$ where $i=1,2$) and the emission probabilities of observed signals given the latent states ($P({X}_{j}|{Z}_{i})$ where $j=1,2,3$). The model captures the relationship between the unobserved states and the observed data over time.
• Figure 4: Inferred emission probabilities and relative risks by news channels. Panels (a)-(c) show emission probabilities associated with ${X}{1}$, with panel (b) detailing $P({X}{1}|{Z}{1})$ for toxic content, panel (c) for insulting content, and panel (a) identifies the probabilities from Figure \ref{['HMMDiagram']} that are emphasized in panels (b) and (c). Notably, $P({X}{1}|{Z}{2}) = 0$ across all channels (indicated by a dotted arrow in panels (a), (d), and (g)), suggesting that conversations do not conclude in state ${Z}{2}$. This, combined with $P({X}{1}|{Z}{1}) > 0$, identifies state ${Z}{1}$ as the likely terminal state, while state ${Z}{2}$ corresponds to earlier stages in a conversation. Panels (d)-(f) display relative risk findings for ${X}{2}$, with non-toxic posts (panel (e)) and non-insulting posts (panel (f)). Here, $RR{{X}{2}} < 1$ across all channels, suggesting that non-toxic or non-insulting activity is more common in state ${Z}{2}$. Panels (g)-(i) summarize the relative risk findings for ${X}{3}$, with toxic posts in panel (h) and insulting posts in panel (i). In almost all channels (with CNN as an exception), $RR{{X}{3}} > 1$, indicating that toxic or insulting posts are more likely when a conversation is in state ${Z}{1}$. This effect is most pronounced for Fox News (a right-leaning channel), while among left-leaning channels, ABC News exhibits the highest relative risk for ${X}_{3}$.
• Figure 5: Inference of hSBM to video descriptions. Results of clustering 19,365 video descriptions and 26,041 words, forming a network with 576,092 edges using the hSBM approach. At the highest hierarchical level, the model separates word nodes from video description nodes, reflecting its bipartite structure. At the fourth hierarchical level, it categorizes words into 10 topics, including two functional word topics (0 and 4), and identifies 9 distinct video clusters.