Research summary. This article summarizes the study AI-Mediated Communication Can Steer Collective Opinion, by the Oxford Internet Institute and the Hasso Plattner Institute, based on the publication by the OII.
Research by the Oxford Internet Institute (OII) and the Hasso Plattner Institute concluded that artificial intelligence tools can alter the direction of posts about controversial topics, even when instructed to preserve the original meaning. The study, titled AI-Mediated Communication Can Steer Collective Opinion, was accepted for presentation at workshops of the ICML 2026 conference, in Seoul, South Korea.
The researchers instructed language models from different providers to transform texts written by people into social media posts. They then analyzed shifts in positions and used mathematical modeling and simulations based on data from real social networks on X and Facebook. The results indicated that small individual changes can accumulate over millions of interactions and gradually shift opinions in online communities.
The research also identified similar trends across different systems, with changes favorable to some positions on gun control, marijuana legalization and feminism, and unfavorable to others, such as atheism and the death penalty. In a simulation of X’s “Explain this post” feature, the researchers observed greater support for posts opposing abortion rights than for posts supporting them, a result associated with a specific instruction for the model to challenge prevailing narratives when necessary.
The study indicates that implementation decisions made by platforms can influence the direction and intensity of this effect, in addition to the behavior of the model itself. The possibility of gradual and barely perceptible persuasion is also highlighted by Dora Kaufman in Demystifying Artificial Intelligence: “These persuasion processes occur at invisible levels, without users’ knowledge or consent. Users are unaware of the potential and scope of artificial intelligence algorithm predictions; at a more advanced level, these predictions involve personality, emotions, sexual and political orientation — in other words, a set of information that the user theoretically did not intend to reveal.” The research argues that AI-mediated communication creates additional challenges for transparency, accountability and regulation, which have not yet been directly addressed by frameworks focused on systemic risks, harmful content, discrimination and threats to democratic processes.
Information from oii.ox.ac.uk.
This post was summarized from the original publication using artificial intelligence, with human review by Guilherme Goulart.