New research from the Oxford Internet Institute and the Hasso Plattner Institute reveals that artificial intelligence tools used to generate or edit social media content can systematically alter public discourse. Published on arXiv and accepted for presentation at the ICML 2026 conference in Seoul, South Korea, the study demonstrates how large language models consistently shift the sentiment of user posts on contested issues, even when explicitly programmed to maintain the original viewpoint. Through mathematical modeling and simulations based on actual X and Facebook network data, researchers showed that these micro-adjustments compound across millions of interactions, gradually steering broader public opinion. A controlled test of X’s Grok Explain this post feature on abortion-related content highlighted the mechanism behind this influence. The model exhibited a measurable pro-life bias, which traced directly to a single prompt directive instructing it to challenge mainstream narratives. The experiment confirmed that minor platform-level instructions can disproportionately shape how AI contextualizes and amplifies online content, turning drafting assistance into a scalable opinion-shaping tool. The findings position AI-mediated communication as a subtle influence mechanism that current technology regulations fail to address. While the European Union’s AI Act and Digital Services Act target systemic risks, harmful material, and algorithmic discrimination, they do not explicitly cover the gradual opinion alteration inherent in AI-assisted social media drafting. Senior author Sandra Wachter, professor of technology and regulation at Oxford, emphasized that legal frameworks have not yet adapted to this new vector of discourse control. The research provides a technical and policy roadmap for understanding how automated language tools reshape digital public spheres, underscoring the urgent need for governance standards that account for the cumulative persuasive power of AI-generated network activity.