TECH

AI-powered social media can subtly manipulate opinion at scale, new study finds

The study found that large language models (LLMs) consistently changed the direction of social media posts on contested topics, even when explicitly instructed to preserve the original meaning. The researchers also show, through simulations of real-world social networks, how these small changes could accumulate across millions of interactions and gradually influence broader public opinion. 

The findings raise questions about the growing use of AI-powered writing tools on social media platforms and suggest that AI-mediated communication could become a powerful new mechanism for influencing public discourse.  

The study, AI-Mediated Communication Can Steer Collective Opinion, authored by Dr. Stratis Tsirtsis, Kai Rawal, Professor Chris Russell, Professor Brent Mittelstadt and Professor Sandra Wachter, has been accepted for presentation at the AI4Good and Technical AI Governance Research workshops at the International Conference on Machine Learning (ICML 2026) in Seoul, South Korea. 

Key findings...AI writing and editing tools can introduce bias into social media posts. Large language models (LLMs) systematically altered the direction of users' messages on contested topics, even when instructed to preserve the original meaning.  

Biases were similar across different AI systems. Multiple models tended to nudge posts in similar directions, favouring some positions such as gun control, marijuana legalisation and feminism, while pushing against others such as atheism and the death penalty.  

Small changes in individual posts can influence public opinion over time. Simulations using real-world social network data showed that subtle biases introduced into posts can accumulate and gradually shift opinions across online communities.  

AI-assisted communication creates a new route for influencing public discourse. The researchers argue that AI systems embedded in social media platforms can shape how opinions spread online, creating new challenges for transparency, accountability and regulation.  

The source of bias is not just the AI model itself. The study shows that specific implementation decisions made by platforms can significantly affect the direction and magnitude of AI-generated influence. 

Methods...The researchers instructed large language models (LLMs) from different providers to transform human-written texts on contested topics into improved social media posts. They then analysed whether the AI-generated versions systematically changed the position expressed in the original posts. Next, they used mathematical modelling and computer simulations based on real social network data from X and Facebook to examine how these small changes could spread through online networks and affect broader public opinion over time. 

Example: Grok’s “Explain this post” feature 

By recreating and testing X’s “Explain this post” feature, with a focus on abortion-related posts, researchers found that Grok was more supportive of pro-life posts than pro-choice posts. By removing X’s instructions one by one, they traced this imbalance back to a single instruction telling Grok to “challenge mainstream narratives if necessary”. This experiment illustrates how targeted, easy-to-implement platform interventions can shape the way AI systems influence public discourse online. 

Implications for regulation...The research highlights AI-mediated communication as an emerging form of influence that existing regulatory frameworks do not yet cover. While initiatives such as the EU AI Act and Digital Services Act focus on systemic risks, harmful content, discrimination, and threats to democratic processes, they do not directly address the more subtle ways AI can shape opinions through drafting, editing, or contextualising online content. 

“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,” says senior author Sandra Wachter, Professor of Technology and Regulation at the Oxford Internet University of Oxford Institute, 

University of Oxford