Can language models stand in for people in social measurement?

I build survey instruments that exploit where they can, and detection methods for where they fail. My work uses large language models to measure public opinion from social media and to probe the limits of synthetic respondents.

PhD candidate in Survey and Data Science, University of Michigan Institute for Social Research, advised by Frederick G. Conrad.

Research program

One question, five threads

i.

Digital twins of survey respondents

When persona-conditioned language models can predict an individual's survey responses, and the mechanisms behind the cases where they cannot.

ii.

Validity of social simulation

Whether populations of language-model agents reproduce social processes, or only their surface outcomes. The population-scale version of the twins question.

iii.

Who social media speaks for

Demographic inference, stance measurement, and platform ecology, toward a dissertation on the conditional representativeness of online expression.

iv.

Language models as survey instruments

AI interviewers and LLM-assisted coding of open responses: building and evaluating the tools, not just studying them.

v.

Structuring text at scale

Efficient ranking and embedding-space summarization for making sense of large social-media corpora.

See how the threads connect →

Selected papers

Recent work

Preprint · 2026 Persona-Based Simulation of Human Opinion at Population Scale. Li, M. & Conrad, F. G. arXiv
Proc. ICWSM · 2026 Who Talks to Whom: Quantifying Echo Chamber Effects in Emerging Social Media Platforms. Li, M. & Chen, X. DOI
Preprint · 2024 Advancing Annotation of Stance in Social Media Posts: A Comparative Analysis of Large Language Models and Crowd Sourcing. Li, M. & Conrad, F. arXiv

All publications →