Working on now
AI for a better
democratic society.
Current projects on using AI to improve society—and to help people take part in democratic life.
A central research program at SAIL · Work in progress
Learning to disagree, productively.
Can AI help people build the skills to navigate political disagreement—and use those skills in everyday life?
Our team is developing a research program around a simple premise: handling disagreement can be learned. We are studying how to assess people's strengths, offer focused training, and provide practice conversations with AI partners.
We focus on five abilities: participating in difficult conversations, expressing a view, giving reasons, listening, and reconsidering one's own perspective. The goal is to help people become more capable conversation partners beyond the training itself.
With Lisa P. Argyle, Joshua R. Gubler, Alex Lyman, and David Wingate.
More about SAIL ↗Work in progress · Political communication
What does trying to persuade do to the speaker?
Political conversation can change the person doing the talking as well as the person listening. With Christopher F. Karpowitz and Lisa P. Argyle, I am examining how the topic and purpose of speech shape speakers' experiences and attitudes.
Participants record videos about political or nonpolitical topics, with instructions to persuade or offer a balanced account. We examine what they say, how they speak, and how they feel afterward. This project asks what kinds of speaking experiences leave people more open to other perspectives.
Work in progress · AI & causal inference
Can AI simulate an experiment—or just a correlation?
When a language model produces responses that look human, does it also reproduce what an experiment would tell us? We are investigating how well AI-generated data distinguish causal effects from observational associations.
We compare simulations with human experimental data and test controlled scenarios, including the difference between assignment to a treatment and actually receiving it. The project examines how results depend on prompting, model choices, and the groups being represented. Our aim is to clarify when these tools can support credible inference and where they fall short.
These descriptions summarize ongoing research. Working manuscripts are not shared here.