AI writing assistants subtly shift opinions in social media posts, study finds

A new study from the Oxford Internet Institute and the Hasso Plattner Institute reveals that AI writing tools consistently alter the sentiment of social media posts, even when instructed to maintain the original meaning. These subtle shifts in framing on contested topics can accumulate across millions of interactions, potentially steering public discourse in specific directions. For the social media sector, this discovery highlights a hidden layer of algorithmic influence that current regulatory frameworks are not yet equipped to address.

Researchers from the University of Oxford and the University of Potsdam tested various large language models to determine if AI-driven editing preserves a user's original intent. The study found that AI tools systematically nudged posts toward specific ideological positions on sensitive issues like gun control, feminism, and marijuana legalization, while pushing against topics such as the death penalty and atheism. This directional pressure occurs even when users explicitly command the AI to keep the original meaning intact, suggesting a built-in bias within the models rather than random error.

The implications for social media platforms are significant, as mathematical simulations using data from X and Facebook demonstrated that these minor per-post biases can snowball into large-scale opinion shifts within online communities. The research specifically examined X’s “Explain this post” feature powered by Grok, finding it favored pro-life perspectives over pro-choice ones. This specific bias was traced back to a single system instruction telling the AI to “challenge mainstream narratives if necessary,” illustrating how platform-level configuration choices can quietly shape the narrative of millions of user interactions.

Senior author Sandra Wachter, a professor at the Oxford Internet Institute, warns that this AI-mediated communication represents a new form of influence that current laws like the EU AI Act and the Digital Services Act fail to regulate. While these frameworks focus on systemic risks and harmful content, they do not account for the subtle ways AI assistants edit or contextualize everyday speech. As these tools become a standard part of the social media experience, the study raises critical questions for the industry regarding whether the opinions expressed online belong to the human users or the underlying models.