ARTICLE AD BOX
AI writing and editing tools on social media do more than just fix your grammar or polish your posts. According to a new study from the Oxford Internet Institute and the Hasso Plattner Institute, they actually influence what people think—little by little—by nudging how users frame their opinions.
The research, called “AI-Mediated Communication Can Steer Collective Opinion,” will be presented at ICML 2026 in Seoul.
Testing Popular Language Models
To dig into this, the team put four popular language models—Llama 3.1, Gemma 3, Mistral, and Qwen—to the test. They ran these models on real social media posts about thirteen hot-button topics, like gun control, abortion, marijuana legalization, and the death penalty.
They told the AIs not to change the message. Still, the researchers noticed that the bots regularly shifted tone, emphasis, or word choice in a particular direction. The study found the models leaned toward supporting things like gun control, marijuana legalization, and feminism. On the other hand, they tended to pull away from supporting atheism or the death penalty.
Individually, each change was subtle. But you know how it goes—a million little nudges add up.
“These systems don’t have to rewrite your post to make a difference,” said the lead authors. “When you shift tone or word choice in the same direction millions of times, people start to see the ‘consensus’ as changed.”
Testing Grok’s Bias
The team also checked out Grok, xAI’s chatbot, by simulating X’s “Explain this post” feature.
After analyzing 78 abortion-related posts, they noticed Grok’s explanations leaned more toward pro-life content than pro-choice. The researchers traced this back to a system prompt that tells Grok to “challenge mainstream narratives if necessary.” They think this bit of guidance interacts with Grok’s training data and tips the scales.
Small Edits, Big Results
To understand the wider effect, the team ran mathematical models and computer simulations using real social network data from X and Facebook. Their results make one thing clear: millions of these subtle AI edits can slowly shift what whole communities believe or perceive online.
Importantly, the researchers aren’t pointing fingers at one company or bad actor. This isn’t overt manipulation—it’s a snowball effect from loads of tiny edits over time.
Regulation Hasn’t Caught Up
The study points out a gap in current AI regulation. Existing laws, like the EU’s AI Act and the Digital Services Act, tend to focus on safety, transparency, or moderating content. None of them zero in on how AI writing and editing tools themselves shape what people think.
“AI-assisted communication is a new kind of influence,” said Professor Sandra Wachter, one of the lead researchers. “Current law just doesn’t catch it all.”
What Should Policymakers Do?
The team put forward a few solutions to make things fairer and more transparent. These include:
- Letting users know whenever AI edits or suggests changes.
- Opening system prompts for review by independent auditors.
- Carrying out regular bias audits on these AI writing features.
- Giving users full control to turn AI help on or off.
Why This Matters
For journalists, content creators, and social platforms, this research rings alarm bells. AI writing assistants can make things clearer, catch harmful or hateful language, and help people communicate in languages that aren’t their native tongue. But those perks come with a real risk: people’s views can be shaped, unintentionally, on big issues.
With AI now part of so many online conversations, the study argues that transparency and user choice aren’t just nice to have—they’re essential for people to keep trusting what they read and share.
Researchers, policymakers, and tech companies will dig deeper into these findings at ICML 2026, figuring out what they mean for the future of conversation and opinion on the internet.
English (US) ·