From Toxicity to Conformity: Adaptive user behavior to social norms in Telegram communities

TL;DR

The study addresses how online toxicity reflects local normative environments by analyzing six large Telegram datasets (over 522 million messages) across languages. It employs Perspective API toxicity scores, logarithmic binning for stable correlations, and per-user linear regressions to derive a conformity index $\\theta$ that classifies users into conformist, anti-conformist, independent, or zen categories, with a threshold $\\tau$ and angle interpretation. Key findings show a strong global association between chat and user toxicity, a majority of conformist users who adapt to local norms, and increasing conformity with greater platform exposure, consistent across datasets and languages. The work highlights that context and social influence drive toxicity patterns, offering norm-aware implications for moderation and platform design while noting Telegram-specific limitations and classifier biases as avenues for future research.

Abstract

Toxic and antisocial user behavior on social media platforms has received considerable scholarly attention due to its detrimental effects on society. This study takes a holistic perspective on the phenomenon of online toxicity by investigating the impact of local community norms on toxic expression. By using six large-scale datasets, comprising over 500 million Telegram messages collected between 2015 and 2024, we analyze toxic user behavior across multiple chats and languages. We introduce a methodological framework that models user adaptation through a conformity index, capturing conformist, anti-conformist, and independent behavioral tendencies. Our findings show that most users tend to conform to local normative environments, adjusting their toxicity to match the toxicity levels of the chats in which they participate. These patterns are consistent across datasets and languages, suggesting that community norms and social influence play a decisive role in shaping user behavior online. Furthermore, we demonstrate that exposure to these norms, in terms of increased user participation in chats, is associated with a stronger tendency toward conformity with the surrounding social contexts. Collectively, these findings contribute to a deeper understanding of toxic online behavior and highlight the importance of contextualized approaches to content moderation.

Figures (5)

• Figure 1: Complementary cumulative distribution function (CCDF) of chat toxicity across datasets. All datasets, despite linguistic and contextual differences, display similar tail behavior indicating the presence of both low-toxicity and highly toxic communities, providing variability for cross-environment behavioral analysis.
• Figure 2: Correlation between user toxicity and chat toxicity. Panel (a) shows that, across all datasets, user toxicity increases with the toxicity level of the chat they participate in. The same pattern holds in Panel (b) when removing the user's own contributions (Leave-One-Out), confirming that the effect reflects genuine environmental influence. Shaded areas represent 95% confidence intervals. Correlation values are reported in Table \ref{['tab:combined_corr']}.
• Figure 3: Change in user toxicity as a function of environmental shifts. For every user active in at least two chats, the plot shows how changes in community toxicity correspond to changes in the user's own toxicity. All datasets display a clear positive trend: users become increasingly toxic when moving into more toxic environments. This dynamic evidence shows that user behavior is shaped by contextual norms rather than static personal tendencies.
• Figure 4: Individual patterns of conformity to community norms: Panel (a) shows two examples of fitted user trends, showing how individual user-specific slopes capture different adaptive behavior to chat environments. Panel (b) shows the distribution of these behavioral types across datasets, where the majority of users appear to be conformists, with independents and anti-conformists representing smaller shares. This pattern remains consistent when tested across different slope thresholds (Figure \ref{['fig:thresh-origin-label']}). Panel (c) shows how these proportions change when increasing the minimum number of messages required for inclusion, revealing that higher user activity and engagement reduces independence and reinforces conformity, as increased exposure to community norms reinforces norm alignment.
• Figure 5: Sensitivity analysis of user classification as a function of the slope angle threshold $\tau$. The plot shows how the proportions of conformist, independent, and anti-conformist users vary as the threshold used to define independence is adjusted from $2^\circ$ to $30^\circ$. Users are classified as independent when $|\theta| < \tau$, where $\theta$ is the regression slope expressed in angular degrees. Despite changes in the threshold, the ratio between conformists and anti-conformists remains approximately stable, indicating that the predominance of conformist behavior is robust across a broad range of cutoff values.