A growing body of academic research and practitioner evidence shows that, under the right conditions, political conversations can reduce political animosity, de-escalate conflict, and facilitate compromise. However, not all difficult political conversations are so idyllic: many are heated, combative, and create feelings of stress, frustration, or anger. Because many people are concerned about such negative impacts on themselves and their relationships, they avoid engaging at all in political conversations with people they disagree with. Even when they do choose to participate in these discussions, people generally lack the civic skills for productive engagement in divisive, disagreement-based, or conflictual interactions.

We propose that productive disagreement is a skill that can be improved with instruction and practice. At the same time, existing programs that can teach this skill are severely resource constrained and face limits on their scalability. In this project, we suggest that generative AI tools can step into this space to both train people to develop these disagreement skills and test transfer from this learning into conversations with others. Specifically, we develop a set of AI agents tailored to provide real-time coaching and training to participants about engaging productively in political discussion using evidence-based best practices drawn from a variety of fields. Post-training, we also deploy a separate AI agent to give people practice engaging in potentially heated discussions without posing any risk to their real-life relationships and as an evaluation tool. We then evaluate whether practice and coaching improve people’s confidence and ability to engage productively in divisive political conversations. In this paper, we introduce our framework and approach to this use of AI and provide empirical demonstrations of its value.