An Oversight Board study says top AI models are stifling political speech

The Oversight Board found that major AI models are more likely to refuse generating politically critical content about repressive governments than about free-speech-protected ones, indicating a form of indirect censorship. This study tested models from Anthropic, DeepSeek, Google, Meta, OpenAI, and xAI, revealing a refusal rate over twice as high for restrictive countries. The board warns this censorship by proxy could affect multiple apps relying on these foundation models and calls for greater