What Makes Constructive Dialogue Work and Fail

Online political discussions are increasingly uncivil and toxic characterized by partisan animosity and polarization. Yet, we still know little about what contributes to constructive online dialogue, and what explains why it often breaks down. Using a large language model in terms of ChatGPT that we integrate into survey experiments in the United States and the United Kingdom (n = 3,307), we study the micro-foundations of political rhetoric and how they affect the quality and content of democratic dialogue as well as broader outcomes. We ask respondents to express an opinion on an issue that matters to them and present them with a counter-argument generated by the large language model that is tailored to their issue position and randomly varies along four treatment dimensions: tone, justification, willingness to compromise, and partisanship. We find that counter-arguments that use evidence-based justifications and signal a willingness to compromise lead to more high-quality responses, foster replies that use sound justifications themselves, and make people more willing to compromise and open to different beliefs. A disrespectful tone, on the other hand, has a substantially negative effect on constructive dialogue, as it results in more disrespectful replies from participants, a lower willingness to compromise, and overall lower quality replies. It also reduces openness to other beliefs. A likely explanation for this is that the style of argumentation affects perceptions of the sender, which we show. We find no effects on political engagement, attitude moderation, or polarization, except that interacting with an out-partisan increases affective polarization slightly. Our research serves to untangle the repercussions of online discussions, bearing significant implications for comprehending the dynamics of democratic dialogue.