Meta Oversight Board finds AI models less likely to criticize repressive governments

A Meta Oversight Board study finds that leading AI models from labs like OpenAI and Google are less likely to criticize repressive governments. The models refused 34% of requests for politically critical content about restrictive regimes, compared to 14% for permissive ones. The board, which operates independently of Meta, urged AI firms to conduct human rights analyses and increase transparency in their training and evaluation processes.

Meta Oversight Board AI study

• ▪The Meta Oversight Board is funded by Meta but operates independently.
• ▪The Meta Oversight Board conducted its first study on large language models, testing models from labs including Anthropic, OpenAI, Google, Meta, and DeepSeek.
• ▪The study categorized 10 jurisdictions as 'permissive' or 'restrictive' using rankings from the NGO Freedom House.

Bias toward repressive governments

• ▪The board found evidence of models explaining they were following explicit rules that did not exist and were not evenly applied.
• ▪A study by Meta's Oversight Board found that leading AI models are less likely to criticize governments known for restricting free speech.
• ▪The study showed that AI services were echoing the rules of countries that restrict speech.

AI model refusal rates

• ▪AI models refused 34% of requests for politically critical content about 'restrictive' jurisdictions like China and Saudi Arabia.
• ▪The refusal rate for politically critical content about 'permissive' jurisdictions, which lack or do not enforce such laws, was 14%.

Recommendations and Governance

• ▪On July 15, 2026, Google DeepMind CEO Demis Hassabis called for a U.S.-led AI watchdog to screen advanced models globally before deployment.
• ▪The Meta Oversight Board urged AI companies to conduct systematic human rights analyses.
• ▪The board also asked for greater transparency in the training and evaluation processes of AI models.

2 sources