Institute for Small Business Economics

Disentangling interaction and bias effects in opinion dynamics of large language models

Abstract Large Language Models are increasingly used to simulate human opinion dynamics, yet the effect of genuine interaction is often obscured by systematic biases. We develop a Bayesian framework to disentangle and quantify three such biases: (i) A topic bias toward the LLM’s...