Daniel R. Cowen

Online political discussions often turn sour once a toxic comment is made, with toxic comments triggering even more toxic responses. This can negatively impact an individual's mental health and lead to hostile perceptions of people holding different views. In this paper, we aimed to understand whether different constructive characteristics of a contribution to a discussion can stop a toxic cascade. Using Perspectives API, we quantify toxicity and seven constructive dimensions of comments left on a political Reddit forum. We focus on five-comment chains in which toxicity escalates and examine whether a constructive intervention at the midpoint of the chain can alter its trajectory Using Bayesian beta regressions, we find that that reasoning, respect, and curiosity are associated with lower subsequent toxicity, particularly in already highly toxic threads. In contrast, nuance and compassion can backfire and increase toxicity in those same contexts. These findings demonstrate layered accounts of constructive communication by showing that not all constructive moves are equally effective, and that some can be reinterpreted negatively when conflict is already high. We discuss implications for theories of online deliberation and conflict, and for the design of platform interventions that selectively amplify reasoning and respect in toxic political discussions.

This paper studies the dynamics of toxicity in online political discussion on a political Reddit forum, where users explicitly identify their political affiliation for all to see. By analysing a dataset of over 2 million comments, we use social identity theory as a tool to understand how group membership, prior interaction and identity salience contributes to the prevalence of incivility in comment and response discussion chains. Our findings reveal how toxicity online very often leads to subsequent hostility, potentially leading to a cascade of incivility. In addition, we show that as ideological distance increases between two individuals, incivility is more likely, particularly when prior toxicity had already taken place. Our study also confirms that closer to election time, there is slightly more intergroup hostility than further away, highlighting the role of identity salience in cross partisan discussions. This work underscores the importance of understanding the mechanisms behind online toxicity to foster healthier online political discourse, as well as contributes to literature of online political discussions, social identity on the internet and the use of largescale online deliberation data to understand cross-group interactions.