A new Meta Oversight Board study finds that leading AI chatbots are more likely to refuse prompts criticizing leaders in countries with restrictive speech laws than those in open societies, raising concerns that language models may inadvertently propagate government limits on expression across borders. In tests of 10 commercial systems from companies including Meta, Anthropic and OpenAI, models often produced critical content about figures such as President Donald Trump or the U.K.’s King Charles III but declined when asked to target leaders in China, Saudi Arabia, Thailand, Turkey and Cambodia.
The board said the pattern could stem from biases baked into training data and developers’ legal-risk calculations, and warned that without human-rights due diligence, AI infrastructure could extend illegitimate speech restrictions globally. A separate academic study published in Nature reported that U.S.-built models gave more state-aligned answers in non-English languages, underscoring vulnerabilities in multilingual settings. Companies contacted by the AP did not immediately respond. The findings land as governments seek AI guardrails, including a Trump administration initiative focused on national-security risks from advanced systems.
Related articles:
— Constitutional AI: Harmlessness from AI Feedback
— Political content policy for Google Ads
