AI Mediation Steers Collective Opinion Formation Online
A recent study published on arXiv (2605.16245) indicates that large language models (LLMs) can influence group opinions during human interactions on social media. The researchers discovered that various well-known LLMs introduce biases when modifying human-generated content on contentious issues, such as promoting views supporting gun control and opposing atheism. By integrating empirical research with a mathematical framework for opinion dynamics, the study illustrates how AI systems, when situated between users in a social network, can alter and shape collective opinion development. This research builds on earlier findings regarding AI's effect on individual beliefs, emphasizing a previously overlooked danger of AI's role in online discussions.
Key facts
- Study published on arXiv with ID 2605.16245
- LLMs polish posts on LinkedIn and provide context on X
- Empirical analysis shows directional biases in LLM edits
- Biases favor gun control and oppose atheism
- Mathematical model introduced for AI-mediated opinion dynamics
- AI sits between users on a social network transforming messages
- Prior work focused on individual opinion, not collective
- Study addresses gap in understanding AI's influence on group opinion
Entities
Institutions
- arXiv
- X