My research investigates and designs human–AI social systems that strengthen rather than undermine society.
Artificial intelligence is rapidly integrating into the social processes through which people communicate, coordinate, and cooperate. Within these human–AI social systems, AI shapes expectations between individuals, social norms across groups, and institutional governance. To understand and engineer these emerging dynamics, my lab combines behavioral experiments, multi-agent simulations, and computational social science.
AI-Mediated Communication
People build relationships, communities, and shared understanding through communication. As large language models become part of everyday conversations, AI increasingly participates in the construction of social reality.
My lab aims to establish design principles for AI that strengthen—not fragment—human relationships by examining the social consequences of AI-mediated communication. In one stream of work, we show that AI personalized only to individual users can reinforce self-expression while reducing cooperation across opinion differences. In contrast, Relational AI, which also considers social relationships, encourages more receptive communication and helps people cooperate across disagreement. We further investigate how these individual interactions scale up to reshape long-term social norms and community structure, leading to either polarization or social cohesion.
We also study communication for collective decision-making under distributed information. Groups can make better decisions than individuals because different members hold complementary insights. In practice, however, discussions often fail because individuals pursue premature consensus instead of sharing unique perspectives (“shared information bias”). Our research shows that state-of-the-art AI agents exhibit similar communication failures. Building on these findings, we explore communication protocols that help people and AI share and integrate distributed information more effectively, enabling superior collective decisions.
Coordination & Cooperation with AI
Many of society’s most pressing challenges are collective action problems in which individual incentives conflict with collective welfare (“social dilemmas”). We explore AI designs grounded in principles fundamentally distinct from those focused on individual performance, efficiency, or convenience.
My lab examines how AI agents can improve coordination and cooperation by reshaping social networks and interactions. Specifically, we have shown that introducing small amounts of behavioral noise can help groups escape coordination failures, and strategically restructuring social connections can improve cooperation and collective welfare across an entire network. These findings demonstrate that designing AI for collective action requires approaches tailored to group dynamics rather than individual usability.
Another line of research addresses shared expectations. People coordinate and cooperate more effectively when they share expectations about one another’s behavior. My lab investigates how AI predictions can align these expectations to improve collective outcomes through self-reinforcing dynamics (“self-fulfilling prophecy”). At the same time, our work reveals an important tradeoff: as coordination depends more heavily on AI assistance, people may rely less on one another, eroding the reciprocity and social intentions that traditionally bind communities together. We work to navigate this balance to design human–AI systems that strengthen collective action without diminishing human sociality.
Behavioral Experiments & Social Simulations
Understanding human–AI social systems requires methods that link granular causal mechanisms to large-scale social dynamics. My lab develops experimental platforms enabling hundreds of participants—and increasingly, AI agents—to interact in realistic social environments. For instance, our large-scale online social experiments recreate communication and coordination during uncertain disasters, shedding light on how misinformation spreads, social networks adapt, and collective behavior emerges.
To complement human-subject experiments, we construct AI-agent simulations to evaluate targeted interventions for large-scale social challenges. In collaboration with emergency preparedness professionals, we design social simulations with more than 10,000 LLM-powered agents that inform public policy. By bridging behavioral experiments with agent-based simulations, we extend empirically grounded insights to societal-scale scenarios—allowing researchers and policymakers to evaluate interventions far beyond the physical reach of laboratory settings.
