How AI editing of posts can shift collective opinion on social networks

When LLMs mediate human-to-human communication by editing posts, they introduce directional biases—nudging texts toward gun control, away from atheism—that cascade through social networks. The researchers combine empirical analysis (testing multiple LLM families on contested topics) with a mathematical model of opinion dynamics to show how these biases compound across users, ultimately shifting collective positions. An audit of X's "Explain this post" feature reveals pro-life bias in Grok's outputs on abortion, traced to specific design choices. The work demonstrates that platform designers can control these biases but currently do not, with implications for EU regulatory frameworks.