Answers are grounded in this paper and cite their sources.
Large language models (LLMs) are increasingly being considered as tools for computational social science, yet their implications for survey and public-opinion research remain insufficiently understood. This study examines whether and how they can serve as methodological instruments for estimating aggregate public opinion under out-of-sample conditions. Across three studies, commercial and open-source LLMs are evaluated against recent Pew Research Centre surveys. Study 1 shows that GPT-3.5 and GPT-4 can approximate national-level response distributions more accurately than uniform, normal-distribution, and historical-average baselines, with GPT-4 achieving mean absolute errors of approximately 4–8.5%. However, the models also exhibit important threats to validity, including underrepresentation of Republican viewpoints, underestimation of uncertainty, and, in some cases, overly homogeneous response distributions. Study 2 extends the analysis to GPT-4o, GPT-4o-mini, and Llama 3.3, showing that aggregate predictive accuracy generalises across models but that partisan alignment and distributional fidelity are model- and topic-dependent. Study 3 demonstrates that collective-level estimation is more valid than aggregating persona-based individual simulations, which produce substantially higher error and stronger homogeneity bias. Ablation tests further indicate that predictions are relatively robust to role prompts and temperature settings but sensitive to survey-item context. These findings suggest that LLMs may be useful as exploratory and supplementary tools in public-opinion research. However, the responsible use of LLMs requires explicit validation against human data, attention to bias and uncertainty, and careful integration with established principles of survey design.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: