Overview

Healthy democracies depend on civic muscle: the capacity to stay in the conversation when things get hard. It is the everyday foundation of pluralism and tolerance. Skills like asking open curious questions, not jumping to blame, or grounding arguments in lived experiences are what helps us sustain relationships across our political differences. Research points to both the importance of this capacity and its erosion in contemporary life, yet most scalable interventions aimed at strengthening it remain stubbornly didactic. 

Although theoretical frameworks on healthy disagreement have been well documented and empirically validated, for many of us these such theories rarely stick: lectures and trainings tend to be infrequent, facilitator-dependent, and light on behavioral reinforcement. As a result, participants often leave with vocabulary but little practiced skill.

The central component we believe is missing is practice: structured opportunities to employ such theory against adversaries who will make resolving conflict difficult. This project investigates whether generative AI can close this gap. To do so, we create scenarios employing voice-based AI where users engage in difficult conversations against a responsive, emotionally difficult counterpart, and receive targeted feedback on their performance. Through randomized experiments and behavioral coding of real conversations, we aim to gather empirical evidence for a scalable teaching method that durably transfers to human interaction, in turn providing a tangible path to building the civic muscle that a healthy democratic society depends upon.