Can language models stand in for survey respondents?
Survey experiments are a staple of economics, and some researchers now simulate respondents with language models ("silicon samples"). Only 6 of 179 recent survey papers carry a language-model keyword, 5 of them in AI journals. Where do simulated answers match real people, and where would they mislead economic research?
What prompted this#
Since October 2025, 179 papers carry a survey, questionnaire or respondent keyword (117 in economics journals, 37 in finance). Only 6 of them carry a language-model, ChatGPT or generative-AI keyword, and 5 of those are in AI journals. "Survey experiment" appears on 15 papers, 14 of them in economics.
Questions#
- For published survey experiments in economics, do simulated respondents reproduce the direction and size of the treatment effects?
- Do they capture differences between groups (income, age, country), or flatten them?
- Could they help design and size real surveys, even if they cannot replace them?
Possible approach#
Replicate published survey experiments with language-model respondents given the demographics of the original sample, and compare the effects with the published ones.
Next step#
Pick ten published survey experiments in economics, simulate them with two language models, and compare the estimated treatment effects.