AI tools that can steer public opinion fall outside current regulation, Oxford study finds
A new study led by the Oxford Internet Institute (OII) at the University of Oxford, with the Hasso Plattner Institute, has found that AI tools used to write, edit and contextualise social media posts can subtly steer public opinion - a form of influence that current regulation does not yet address.
The researchers found that large language models (LLMs) consistently changed the direction of posts on contested topics, even when told to preserve the original meaning, and that these small shifts could accumulate across online networks to move broader opinion over time. Frameworks such as the EU AI Act and Digital Services Act focus on systemic risks, harmful content and threats to democratic processes, but do not directly address this subtler route of influence.
“Our research points to AI-mediated communication as a new and more subtle way of influencing opinions - one the law has yet to catch up with - and offers food for thought about who, or what, is shaping public discourse.”
The team asked LLMs from several providers to rewrite human-written posts on contested topics as 'improved' versions, then checked whether the position expressed had shifted. Different models nudged posts in similar directions - favouring positions such as gun control, marijuana legalisation and feminism, while pushing against others such as atheism and the death penalty. Using simulations based on real network data from X and Facebook, the researchers showed how these individually minor changes could gradually shift opinion across online communities.
The bias did not come from the AI model alone. Decisions made by platforms about how models are deployed significantly affected the direction and scale of the influence. In one test, the researchers recreated X's 'Explain this post' feature and applied it to abortion-related posts, and found that Grok, the platform's AI model, was more supportive of pro-life posts than pro-choice posts. Removing X's instructions one by one, they traced the imbalance to a single instruction telling Grok to 'challenge mainstream narratives if necessary' - showing how a small platform decision can shape the way an AI system influences public discourse, and how far such choices sit beyond current oversight.
The study, 'AI-Mediated Communication Can Steer Collective Opinion', was presented at the AI4Good and Technical AI Governance Research workshop at the International Conference on Machine Learning (ICML 2026) in Seoul.
For more information about this story or republishing this content, please contact [email protected]
