AI-Mediated Communication Can Steer Collective Opinion

Generative artificial intelligence (AI) is increasingly integrated into the online platforms where humans interact and exchange opinions---for example, large language models (LLMs) are used to improve users' posts before publication and generate contextual annotations around the content users share. While prior work has shown that AI can express biased opinions and also shift individuals' opinions during human-AI interactions, less attention has been paid to how it can influence the opinion formation process of populations when mediating human-to-human communication at scale. In this work, we address this gap via a combination of empirical and theoretical analysis. We start by showing empirically that LLMs from multiple popular families inadvertently introduce directional biases when instructed to draft or improve human-written texts covering a variety of politically salient topics. Building upon this observation, we introduce a mathematical model of AI-mediated opinion dynamics on a social network in which users' opinions are transformed by an AI system before reaching other users. By analytically characterizing the equilibrium of this model and performing simulations on real social network data, we show how even small biases introduced by AI in human-to-human communication can get amplified through the network and shift collective opinion. Finally, we show empirically that these biases are easily controllable in practice through simple interventions to the instructions given to an LLM, demonstrating that AI is an additional lever for online platforms to influence opinion formation. We encourage readers to refer to the extended version of the manuscript at https://arxiv.org/abs/2605.16245.