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

This study proposes and validates a methodological framework using Large Language Models (LLMs) as synthetic respondents in indirect electoral surveys, demonstrating that when processed via Network Scale-Up Methods (NSUM), LLM-generated data can effectively replicate aggregate voting patterns observed in real-world indirect surveys across elections in Spain and Chile.

Original authors: Lucia Soria Gonzalez, Milene Gutierrez Santibanez, Sergio Dıaz, Juan Marcos Ramırez, Rosa Lillo, Antonio Fernandez Anta, Jose Aguilar