Synthetic surveys fail timely topics

When asked to act as ‘digital twins’ of real panelists, large language models produced answers that often diverge from human responses, especially on time‑sensitive or factual questions, and results vary by model. Pew’s experiment using Anthropic’s Claude Opus 4.6 shows synthetic polling can misstate certainty, factual knowledge, and the diversity of opinion compared with high‑quality human panels. — If synthetic polls are treated as equivalent to human surveys they can mislead news coverage, political strategy, and public debate about rapidly changing events.