Applications of Large Language Models in Public Opinion Research: From Silicon-Based Sampling to a Practical Guide

This article introduces the companion GitHub code repository for the monograph Large Language Models for Public Opinion Research: A Practical Guide published by Cambridge University Press, maintained by authors such as Ryan Kennedy. It systematically explains how to use large language models (LLMs) for public opinion research, covering the complete workflow including API interaction, prompt engineering, and silicon-based sampling (synthetic data generation and validation). The code is written in R language, and it also discusses methodological controversies and diverse application scenarios.

Traditional public opinion surveys rely on expensive telephone interviews or online questionnaires, with high sample acquisition costs and long cycles. The rise of LLMs has spawned a new paradigm of "silicon-based sampling"—allowing AI to simulate human respondents to generate synthetic data, promising low-cost and rapid generation of large-scale data, but also sparking profound controversies about validity and representativeness.

The project consists of two core chapters: Chapter 1 introduces the basics of the Transformer architecture, OpenAI API calls (simplified by the ellmer package), the CREATE prompt framework (Context/Role/Examples/Audience/Tone/Extras), and control of output parameters (such as temperature); Chapter 2 explains the silicon-based sampling process: demographic feature extraction → prompt template construction → GPT-5-mini generation of simulated responses → structured storage of results.

Using a subset of the 2021 Cooperative Congressional Election Study (CCES) as benchmark data, multi-dimensional validation is adopted: distribution comparison (differences in key variable distributions), cross-tabulation analysis (variable association patterns), statistical inference (model coefficient comparison); meanwhile, an Ollama local model option is provided to address API cost and privacy concerns.

Supporting views: high cost-effectiveness, precise experimental control, privacy protection; skeptical voices: training data bias, questionable authenticity of responses, external validity to be verified. The authors do not avoid controversies; they demonstrate methods for evaluating the quality of synthetic data through strict validation processes and maintain a cautious attitude.

Social science researchers: learn AI-assisted survey experiments; data scientists: master best practices in prompt engineering; policy analysts: obtain public opinions at low cost; contributors to methodological literature: promote the establishment of field norms.

This project is a milestone in the application of LLMs to social sciences, providing runnable code and a methodological framework (responsibly integrating AI tools, balancing innovation and rigor); as LLM capabilities improve, silicon-based sampling may become a routine tool in public opinion research, and this project provides the technical foundation and ethical guidelines for this transition.