Meta Oversight Board Finds AI Models Censor Restrictive Governments

The Meta Oversight Board released a study finding that large language models frequently refuse to generate political criticism of restrictive governments while complying with similar requests regarding permissive ones. In an evaluation of 10 commercial models from developers including OpenAI, Meta, Google, Anthropic, DeepSeek, and xAI, the board found that AI services refused 34% of requests concerning restrictive jurisdictions—such as China, Saudi Arabia, Thailand, Cambodia, and Turkey—compared to only 14% for permissive regions.

The board characterized this phenomenon as censorship-by-proxy, noting that models often cited foreign laws or invented non-existent policies to justify refusals, even when queried from countries where those laws do not apply. Specific examples included Google's Gemini 3 Pro citing lèse-majesté laws regarding Thailand, and DeepSeek-V3 returning favorable opinions on China 100% of the time. Conversely, models such as Grok 4 Fast exhibited minimal refusal rates.

These findings align with a May study in the journal Nature suggesting U.S.-built models are vulnerable to foreign controls when trained on non-English data. The board urged AI companies to implement human rights due diligence, increase transparency in training, and notify users when legal restrictions shape a refusal. While the board has no binding authority over these providers, the findings emerge as the Trump administration increases oversight of national security risks associated with advanced AI. Separately, Google DeepMind CEO Demis Hassabis called for a U.S.-led watchdog to screen advanced models globally before deployment.