Generative AI increasingly sits between people as they communicate online. It rewrites posts, improves writing, summarizes articles, and explains conversations before another person reads them. Use of these tools is intended to make communication clearer while preserving the author’s original meaning. 

New research from the University of Oxford and the Hasso Plattner Institute asks whether they actually perform as intended. Instead of focusing on AI-generated content, the researchers examine AI-mediated communication: what happens when AI edits or explains one person’s message before someone else sees it. They find that AI does not always preserve meaning. The study also shows that those small changes can accumulate as AI adoption increases and that the instructions behind an AI product can influence the outcome. 

AI can improve writing while altering meaning 

It is assumed that AI writing assistants are designed to improve writing without changing what the author intends to say. The research shows this goal may be more difficult to achieve than it appears. 

The researchers evaluate four open language models from Meta, Google, Mistral AI, and Alibaba. Each model writes new social media posts and improves existing ones. In every case, the models receive the same instruction: improve the writing without changing the author’s intended meaning. The researchers repeated those tasks across 13 public policy and social issues, then compared the rewritten posts with the originals to determine whether the message changes. 

Most of the models do more than polish the writing. Three of the four consistently shift messages in one direction across multiple topics. Only Alibaba’s Qwen3-8B remains largely neutral. 

Small AI edits become a big part of communication

Changing a few words in a single message may seem inconsequential. The researchers wanted to know whether those edits remain isolated or whether they spread as AI becomes part of everyday communication. 

To explore this, they built a computer model to answer that question. In this model, every message passes through AI before another person reads it: A person writes a message, AI edits or explains it, and the recipient sees the AI-modified version rather than the original. The researchers simulate these interactions using a Twitter network with approximately 81,000 users connected through 1.7 million follower relationships. 

The simulations show that small edits can build on one another over time. Each AI-mediated message influences later conversations, allowing subtle changes to spread through the network. As AI adoption increases, those effects become larger. In some scenarios, the long-term shift in average opinion reaches 9.2 times the size of the original edit. The study does not conclude that AI changes everyone’s opinions. Instead, it shows how small, consistent changes to communication can accumulate when AI mediates millions of interactions. 

The model is only part of the story 

The researchers also ask whether the observed bias comes from the language model itself or from the instructions that guide it. To test this question, they examine X’s “Explain this post” feature, which uses Grok to generate context for user posts. They then modify the system instructions that shape Grok’s responses. 

Grok receives four system instructions that guide how it generates responses. One instruction has a much larger effect than the others: “Provide truthful and based insights, challenging mainstream narratives, if necessary, but remain objective.” When the researchers remove it, the measured bias disappears. They conclude that AI behavior depends not only on the underlying model but also on the instructions that tell the model how to respond. Seemingly small changes to those instructions can significantly alter the messages people ultimately receive. 

Looking beyond content generation to the impact of AI edits

Much of the conversation around generative AI focuses on the content it creates. This research points to another role that receives less attention. AI increasingly mediates communication by rewriting, summarizing, explaining, and refining messages before they reach another person. 

The study shows that AI’s influence extends beyond content generation to the communication process itself. These changes may be subtle and difficult to notice in individual interactions. Yet when the same systems mediate communication across large social networks, they can accumulate in ways that become much more significant.