Artificial intelligence cannot replace people in opinion polls because AI-generated synthetic respondents give answers that differ significantly from real respondents, with gaps up to 40 percentage points.

Analysis

People may assume AI can instantly generate accurate public‑opinion data because models can process demographic profiles at scale and produce answers in seconds. The study shows that models rely on training data that may be outdated and tend to apply generic stereotypes, leading to large systematic errors, especially on fast‑moving issues. This matters because pollsters seeking cheaper methods might adopt synthetic respondents, potentially publishing misleading results that shape media narratives and policy decisions. Until models can reliably reflect nuanced, up‑to‑date human attitudes, traditional surveys remain essential.