AI Polls Flatten Public Views

When large language models are prompted to impersonate real respondents, they tend to eliminate the tail answers and concentrate responses toward common or moderate options, and they overstate certainty on factual questions. This experiment by Pew using Anthropic’s Opus 4.6 on American Trends Panel profiles quantifies that effect across multiple real surveys. — If synthetic polling systematically narrows expressed opinion and inflates confidence, media, campaigns, and policymakers relying on AI‑generated survey data risk misreading minority views and public uncertainty.