AI Roleplay Fails As Poll Substitute

Pew’s prompt shows researchers can make a model ‘play’ a respondent by feeding profile and prior-answer JSON and by instructing the model to mimic human ignorance and biases. Even with careful prompts, synthetic respondents require many ad‑hoc constraints and still risk missing human-like nonresponse, uncertainty, and real-world reporting error. — If synthetic polling is imperfect, media, campaigns, and policymakers must treat AI-generated public-opinion data with caution because it can misrepresent uncertainty and diversity of views.