Oversight Board Finds Major AI Models Skew Political Speech by Government Type

Meta's Oversight Board tested 10 AI models from OpenAI, Google, Anthropic, xAI, and Meta with political criticism prompts. Models were more likely to discourage protesting speech-restrictive governments while encouraging support for speech-permissive ones, amplifying the criticized government's posture rather than applying a consistent standard. Co-chair Paolo Carozza called it "extended censorship by proxy that goes across borders."