You ask an AI to polish your post. It keeps your point, but tilts it. New Oxford research says those tiny tilts can move public opinion.
AI writing tools that tidy up social media posts can quietly shift what they say, and those shifts can spread, according to a new study from the Oxford Internet Institute and Germany’s Hasso Plattner Institute. The Guardian reported the findings.
The nudge
The researchers fed human-written texts on contested topics to large language models and asked them to turn each into a better post. Even when told to keep the original meaning, the models shifted the position of the message.
The tilts were consistent. Across different systems, the AI leaned the same way. It nudged posts towards gun control, marijuana legalisation, and feminism, and against atheism and the death penalty.
Small pushes, big drift
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!
One rewrite barely matters. Millions do. Using real network data from X and Facebook, the team simulated how these little edits ripple through a social graph. Over time, the tiny biases pile up and drag community opinion in one direction.
The model is not the only culprit. Platform choices shape the outcome too. The team rebuilt X’s “Explain this post” feature and pointed it at abortion posts, and Grok leaned pro-life. They traced the tilt to a single line in its instructions, one telling it to “challenge mainstream narratives if necessary.”
Why it matters
AI now sits inside the tools we post with. LinkedIn polishes updates, Grok anchors X, and Google is weaving AI through search, quietly steering what we read and buy. Each rewrite is small, but together they form a new and largely invisible channel for shaping opinion.
The law has not caught up. Platforms already sit in court over the harms of their feeds, and regulators are still fighting older battles like age checks. Europe’s headline rules, the EU AI Act and Digital Services Act, target harmful content, discrimination, and threats to democratic processes, not the quiet reshaping of a sentence.
“Our research points to AI-mediated communication as a new and more subtle way of influencing opinions,” said senior author Sandra Wachter, “one the law has yet to catch up with.” The open question is who, or what, is steering the conversation. And whether anyone can see it happening.
