My research asks how AI can improve society and strengthen democracy, from helping people navigate disagreement to giving social scientists better ways to understand human behavior.
Four directions in my AI research
Simulating people
How closely can a language model reflect a human population?
The study proposes that language models can simulate specific human subpopulations when conditioned on detailed demographic backstories, while emphasizing the need to evaluate algorithmic fidelity carefully.
A public working paper asks what counts as evidence of public opinion when AI can generate survey responses. It emphasizes keeping research anchored in the people it aims to represent.
What can AI help us learn about political persuasion?
Two survey experiments use generative AI to test theories of persuasion. AI-generated counterattitudinal messages changed some attitudes, but customization and interaction were not consistently more persuasive than a generic message.
2025 · Proceedings of the National Academy of Sciences 122 (18): e2412815122
Forthcoming at PNAS
Gender, voice, and political influence.
Real-time conversations with AI agents let us separate voice-based gender cues from rhetoric and discussion topics. This study examines how these features shape political persuasion.
Customization changed some messages without improving persuasion overall in the human study. The paper examines what models actually do when asked to tailor political arguments.
Can AI help people have better political conversations?
Evidence-based AI suggestions improved reported conversation quality, democratic reciprocity, and tone without systematically changing the content of conversations or policy attitudes.
2023 · Proceedings of the National Academy of Sciences 120 (41): e2311627120
Evaluating models
How does alignment affect AI’s usefulness for social science?
The study compares aligned and unaligned open-source language models with human responses. It shows that alignment affects simulated attitudes in predictable ways and that model choice matters for social-science research.
2025 · Sociological Methods & Research 54 (3): 1110–1155
2026 · Chinese Political Science Review
Looking under the hood.
Models varied in how they followed customization instructions. The paper connects evaluation of model behavior with evidence about how human readers respond.
Our Arti-‘fickle’ intelligence article puts scientific inference at the center of model evaluation: validating outputs, documenting where models succeed or fail, and connecting those observations to claims about people.
Can AI help people learn to disagree well? A central research program in our lab studies how AI can support the skills democratic conversation requires: participating, expressing a view, giving reasons, listening, and reconsidering one's own perspective.
We are developing ways to assess these skills, offer focused training, and provide practice with AI conversation partners. A key question is whether that practice carries over into conversations with other people. The program is in development; its effects are still being evaluated.
These questions grow out of my work on extremism, partisan stereotypes, populism, listening, and responsiveness. Across these projects, I study how people understand one another and how institutions respond.
In Speaking Populism, we compare how framing, political issues, and ideology shape the ways people express populist ideas.