Model Choice Skews Synthetic Polls

AI‑generated survey responses vary meaningfully by which large model is used, even when given the same persona data and question scripts. In a direct test, Pew asked Anthropic and OpenAI models to play 'digital twins' of real panelists and found systematic differences from human results and between models. — If media, campaigns, or firms publish AI‑generated poll results without accounting for model variance, public opinion measurement and consequential decisions (e.g., messaging, policy signals) could be distorted.