Meta AI models show deference to repressive regimes

Meta Oversight Board found that leading AI models demonstrate reduced willingness to criticize repressive regimes. The board's analysis examined responses across geopolitical topics, noting a pattern of deference when discussing authoritarian governments. This finding emerges as AI providers face increasing scrutiny about model behavior and alignment. The board's work represents one of the first systematic examinations of how AI models handle politically sensitive content at scale. The implications extend beyond Meta to the broader AI industry, as companies grapple with balancing safety measures against potential over-cautiousness in political contexts. The board's findings suggest current model training and filtering approaches may inadvertently suppress legitimate criticism of repressive regimes.