What happened
A Meta Oversight Board study reveals major AI models, including those from Anthropic, Google, Meta, and OpenAI, are significantly more likely to refuse politically critical content about restrictive governments (34% refusal rate) than permissive ones (14%). This behaviour, observed even in US-built models, risks spreading government influence over online speech, particularly when trained on non-English data. The study suggests AI infrastructure may intentionally or unintentionally propagate illegitimate speech restrictions globally.
Why it matters
State influence now extends via AI models, impacting global freedom of expression. Security architects and procurement teams must assess the risk of deploying models embedding foreign censorship, as the mechanism involves training data and developer-implemented mitigation. This follows China's recent curbs on AI companion features, highlighting a pattern of state control over AI outputs.
