Researchers from Oxford and Potsdam examined models from xAI, Meta, Google, Alibaba, and Mistral, finding that AI tools introduced ideological bias when rewriting user drafts on sensitive topics, even when instructed to maintain the original meaning.
In several instances, the systems completely altered the meaning of the original messages. The authors warned that small shifts could amplify over millions of interactions, influencing long-term public opinion, and highlighted a gap in regulation.