Representation Bias in Political Sample Simulations with Large Language Models

This study explores representation biases in simulating political samples using Large Language Models (LLMs), focusing on vote choice and public opinion. We evaluate seven LLMs from diverse cultural backgrounds with data from the American National Election Studies, German Longitudinal Election Study, and Zuobiao Dataset. The study identifies three dimensions of bias: societal and cultural contexts, demographic groups, and political institutions. Results show higher simulation accuracy for vote choice than public opinion, particularly in English-speaking and democratic countries with bipartisan systems. The cultural backgrounds of development teams significantly influence simulation performance. The findings offer insights into addressing biases in AI-driven computational social science.