A synthetic audience is a language model asked to answer as if it were a group of people you cannot easily survey. Drag the dial from a group anyone can poll toward one almost nobody can reach, and watch two things move in opposite directions.
Illustrative figures, not one recorded prompt. The pattern is documented: synthetic panels over-determine segments, inflating between-group gaps two to four times (Chen, Zhu and Zheng, 2026); an expert persona improves how expert the answer sounds while lowering its accuracy (Hu et al., 2026); Artificial Societies reports 86% distribution accuracy against general-population surveys, Evidenza 88% against a benchmark it selected. No peer-reviewed study validates a synthetic expert against held-out real experts of the same kind. Assembled 2026-08-18. A Webflow embed cannot call a model.