Polling Without People: Can Silicon Samples of CASM-Inspired LLM Personae Mimic Human Respondents in Pre-Election Polling?

Abstract

We introduce a framework for simulating pre-election survey responses with large language model (LLM) synthetic personae.Drawing on the Cognitive Aspects of Survey Methodology (CASM) literature, our approach combines synthetic biographies, salient-event memories and conditional news exposure with a staged cognitive process for answering turnout and candidate-choice questions.We apply the framework to a single weekly wave for the 2026 US House elections and produce national and district estimates using multilevel regression and post-stratification.Relative to a demographicsonly synthetic personae baseline, CASM-inspired enrichment shifts estimates towards Democrats, consistent with the direction suggested by contemporary polling.However, the estimated Democratic margin is 18.7 percentage points, compared with 8.7 points in the polling average.Personae randomly assigned to news-based judgement construction produce a larger Democratic margin than retrievers, who are not exposed to news, although uncertainty is substantial.These comparisons show that our enrichment strategy changes synthetic responses in the correct direction, but do not isolate the effects of all components or establish predictive accuracy conclusively.The framework offers a structured approach to studying synthetic survey responses whilst highlighting persistent challenges of leftward bias, uncertainty calibration and benchmarking novel methodologies during an election campaign.