AI-Powered Social Media Platforms Systematically Bias Public Opinion, Study Finds
Researchers from Oxford Internet Institute and Hasso Plattner Institute found that AI language models used to generate or edit social media posts systematically introduce hidden biases that alter the meaning of messages on controversial topics. Simulations showed these subtle changes can accumulate across millions of interactions to gradually shift public opinion.
What happened
Researchers from the Oxford Internet Institute and Hasso Plattner Institute discovered that large language models (LLM) and complex linguistic models (MLC) used to generate, edit or contextualize social media posts systematically modified the meaning of messages on controversial topics, even when explicitly instructed to preserve the original meaning.
The study found that multiple AI systems introduced similar biases across different models, with several tending to influence posts in the same directions—favoring positions such as gun control, marijuana legalization, and feminism, while opposing atheism and the death penalty.
Reported impact
Simulations using real social media data demonstrated that subtle biases introduced in individual posts could accumulate across millions of interactions and gradually shift public opinion in online communities.
Researchers warned that AI-assisted communication could become a new and powerful mechanism to influence public discourse, raising concerns about the growing use of AI-powered writing tools on social media platforms.
Index analysis
The study examined how AI writing and editing tools integrated into social media platforms could introduce systematic biases that shape user-generated content on controversial topics.
The research demonstrated through network simulations how incremental bias introduced by AI systems could compound over time to influence collective opinions at scale.
Governance controls
The research was accepted for presentation at the AI4Good and AI Technical Governance Research workshops at the International Conference on Machine Learning (ICML 2026) in Seoul, South Korea.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.