When AI edits a message, can it shift a whole network’s opinions?

The question: what happens when AI mediates conversation?

Research has shown that language models can express biased opinions and influence an individual in a conversation with AI. The authors focus on a different setting: people communicating with one another through AI tools. A system that rewrites what one person says may also alter what another person receives. Their central question is whether directional changes can affect collective opinion, rather than just individual messages.

Three approaches to the problem

The authors combine empirical and theoretical analyses. First, they test whether language models change the stance of human-written texts when instructed to edit them on contested topics. The abstract reports that models from multiple popular families introduce directional biases, including edits that nudge texts in favor of gun control and against atheism. It does not provide sample sizes, a model list, prompts or measurements, so those details cannot be assessed from the available source material.

Second, the researchers develop a mathematical model in which an AI system sits between users in a social network. The system transforms the opinions users express and perceive. The authors analytically characterize the model’s equilibrium—the state it tends toward—and run simulations using real social-network data. They report that, in the model and simulations, AI-introduced biases can be amplified through the network and shift collective opinion in the direction of the bias.

Third, they audit X’s “Explain this post” feature. The authors report evidence of pro-life bias in Grok’s outputs on abortion-related content, which they say they trace to specific design choices. The abstract does not describe the audit’s size, criteria or those design choices, so their details cannot be evaluated here.

What the findings do—and don’t—show

The authors’ results suggest a possible feedback process: biased edits influence what people say or perceive, and repeated interactions across a network may magnify the initial direction. The network effect is a result of mathematical analysis and simulations, not proof that AI editing has already shifted public opinion in the real world. The abstract reports findings from text-editing tests and the X audit, but does not provide enough methodological detail to independently assess them.

That distinction matters. The abstract does not say how representative the tested texts, topics, models or network data are. Nor does it establish the size of any real-world collective shift. Findings on particular contested topics should not be taken as evidence that every model, editing tool or political issue is biased in the same way.

The authors also investigate whether platforms can control these biases, but the abstract gives no detailed results on that question beyond the reported Grok audit. Their discussion connects the findings to legislative efforts in the European Union; the available source does not specify policy measures they recommend or evaluate.

Why the question matters

When a platform adds an AI layer to communication, its design choices become part of how messages are transformed and perceived. A tool may not need to persuade a user directly to matter: systematic changes to wording or context could build across interactions. That possibility makes transparency and auditing important questions for platforms and policymakers, although this abstract alone cannot establish the scale of any effects.

This article is based on an arXiv preprint, and the available evidence is its abstract rather than the full paper. The claims should be read as findings reported by the authors and model-based implications, not settled evidence of population-wide effects. The work identifies a mechanism worth testing further: whether directional AI mediation measurably changes real conversations and collective opinion over time.