AI Models Censor Criticism of Repressive Governments

Meta's Oversight Board tested 10 commercial AI models from Anthropic, DeepSeek, Google, Meta, OpenAI, and xAI, finding models refused 34% of requests for politically critical material about restrictive jurisdictions versus 14% for permissive ones. The board calls this censorship-by-proxy, as models were queried from an Australian IP where no such speech laws apply.