Use of LLMs for Generating Synthetic ARDs to Make Estimates in Electoral Processes

<title>Abstract</title> <p>Context: Accurately estimating electoral results remains a challenge in political science due to response biases in traditional direct surveys. Indirect survey techniques, which infer voting intentions from respondents’ social networks, have emerged as a promising alternative by reducing response bias and improving scalability. In parallel, large language models (LLMs) have shown a remarkable ability to capture and reproduce social patterns, suggesting their potential use as synthetic respondents in survey-based research. While indirect survey techniques—asking about the voting intentions of acquaintances—mitigate the biases in traditional direct surveys, the emergence of LLMs offers a novel opportunity to simulate these processes using synthetic agents. Objective: This study aims to develop a robust methodological framework to employ LLMs as artificial respondents in indirect electoral surveys, assessing their ability to reflect collective social behavior. Method: To this end, LLMs are prompted to answer indirect survey questions regarding perceived voting intentions within social circles, and the resulting synthetic data are processed using Network Scale-Up Methods (NSUM). Specifically, this research proposes a methodology using specific prompt templates to elicit synthetic responses from an LLM. These responses, focused on perceived voting intentions within social circles, are processed using NSUM. Results: This study provides methodological guidelines and prompt templates that support the integration of LLMs into electoral analysis frameworks. The methodology is applied to two empirical cases: the latest regional elections in Spain and recent national elections in Chile. The study systematically compares different LLM versions, deployment environments, and demographic configurations (age, sex), and contrasts LLM-based estimates with results from real-world indirect surveys. The results show that NSUM-based estimates derived from LLM-generated surveys reproduce key aggregate patterns observed in empirical indirect survey data, although variability across models and settings is observed. Findings: The main finding of this work is that LLMs can infer a collective pattern of electoral behavior, offering a scalable and reproducible complement to traditional indirect surveys, positioning LLMs as a powerful complementary tool for modern political analysis and social pattern estimation. The results demonstrate that the synthetic data generated by LLM can effectively replicate voting behaviors comparable to those obtained through indirect human surveys.</p>