Companies are using large language models to simulate survey respondents and then publish or feed those outputs into media stories as if they were real‑world poll results. These synthetic samples can replicate toplines cheaply but introduce hard‑to‑detect biases and are often reported without disclosure.
— Undisclosed synthetic polling threatens the legitimacy of survey evidence, can mislead journalists and voters, and demands new disclosure and provenance norms for public opinion data.
Joy Li
2026.09.30
85% relevant
The Pew report provides empirical tests showing synthetic samples diverge from real public‑opinion responses on topical questions, knowledge/certainty, and demographic diversity—directly supporting and adding measured nuance to the claim that AI‑generated polls can look like real polls but fail on important dimensions.
Joy Li
2026.09.30
90% relevant
The report provides direct empirical evidence about whether AI‑generated 'synthetic samples' can stand in for human survey respondents. Pew’s digital‑twin experiment (using Anthropic Claude Opus 4.6 and ATP persona profiles) shows key mismatches — especially on uncertainty and factual items — that validate and refine the idea that AI polls can look like polls but fail to reproduce human opinion reliably.
Eli McKown-Dawson
2026.04.11
100% relevant
Aaru (valued at $1 billion) and the Public Sentiment Institute’s practice of boosting 373 real respondents with 114 AI agents, and Axios reporting Aaru findings without noting they were LLM‑generated.