AI summaries in social media improve dialogue but reduce engagement

Generative artificial intelligence agents are becoming increasingly active participants in conversations on social media platforms, yet little is known about how they shape public discussion of social problems. We present two preregistered online experiments testing AI-generated summaries in simulated, interactive social media environments. AI summaries increased the quality of user comments, without systematically increasing toxicity or negative affect. At the same time, AI exposure reduced engagement with conversation threads. AI summaries also increased the semantic similarity between user comments and the AI-generated summaries, suggesting that these systems function as informational anchors that shape discussion. Together, the findings reveal a tradeoff: AI-generated summaries can improve conversation quality while narrowing conversational engagement and channeling how users articulate political arguments. These results speak to growing concerns about how embedded AI systems fundamentally alter platform dynamics and shape public discourse.