A landmark study presented at an international conference finds LLMs nudge posts on contentious topics, compounding biases across platforms and models
A groundbreaking study released on 6 July 2026 has indicated that AI writing and editing tools embedded within social media platforms can introduce subtle but systematic biases into user-generated posts.
Over time, these small shifts accumulate across millions of interactions, gradually reshaping public opinion on polarizing issues. The research was conducted by the Oxford Internet Institute (University of Oxford) and the Hasso Plattner Institute (University of Potsdam).
In their experiments, the team tested four large language models — Llama 3.1, Gemma 3, Mistral and Qwen — using real social media posts about 13 contentious topics such as gun control, abortion, and the death penalty. Even when the AI systems were explicitly instructed to preserve the original meaning of each post, they consistently nudged the content in particular ideological directions.
The models tended to favor positions supporting gun control, marijuana legalization, and feminism, while pushing back against atheism and capital punishment. For example, the phrase “AI might be a useful tool for personalizing the education of students” was transformed by an AI system into “Let’s embrace the potential of AI to personalize learning and revolutionize education for every student.” Although the core idea remained the same, the revised version carried significantly more enthusiasm and persuasive force.
Subtle but powerful psychological manipulation
The researchers had also examined X’s “Explain this post” feature, which is powered by xAI’s chatbot Grok. In 78 posts related to abortion, the scientists discovered that the chatbot had showed more support for pro-life content than for pro-choice content. By systematically removing X’s instructions one at a time, the team traced this imbalance to a single directive telling Grok to “challenge mainstream narratives if necessary.”
Using mathematical modeling and simulations grounded in real data from social networks such as X and Facebook, the peer-reviewed study has demonstrated how these minute individual-level biases can compound and gradually shift collective opinions across entire online communities.
The paper, accepted for presentation at the International Conference on Machine Learning (ICML 2026) in Seoul, underscores a significant regulatory gap. Current frameworks, including the EU AI Act and the Digital Services Act, do not address these subtle yet powerful forms of influence.
According to a senior author of the study, Sandra Wachter, Professor of Technology and Regulation, Oxford Internet Institute: “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.”
