Meta Oversight Board finds top AI models defer to repressive regimes

Meta's Oversight Board released its first evaluation of large language models on Wednesday, finding that AI services from leading labs including Anthropic and OpenAI are much less likely to criticize governments known for restricting free speech, according to Reuters. The study showed that AI models were echoing the rules of countries that restrict speech, raising concerns about the technology's impact on political expression worldwide.reuters+3

A First-of-Its-Kind Evaluation

The report marks the Oversight Board's inaugural foray into evaluating LLMs, expanding its mandate beyond content moderation decisions on Meta ↗ 0.71% platforms. The Board, which typically reviews cases involving Facebook, Instagram, and Threads content standards, found that some of the world's most-used models may be suppressing legitimate political discourse by deferring to the speech restrictions of authoritarian regimes rather than upholding international human rights norms around free expression.oversightboard+3

Research Points to Deeper Structural Biases

The Oversight Board's findings arrive amid a growing body of academic research documenting political biases embedded within AI systems. A study published in Nature's npj Artificial Intelligence earlier this year by researchers including those connected to European institutions found that "large language models reflect the ideology of their creators," after prompting 19 popular LLMs to describe nearly 4,000 politically relevant persons.nature+1

Separately, research published in the Proceedings of the National Academy of Sciences found that LLMs "stereotype the moral values of non-Western populations in predictable ways," portraying people in less-WEIRD (Western, educated, industrialized, rich, and democratic) nations as less morally concerned than they actually are. A Stanford study from 2025 also found that users overwhelmingly perceive popular LLMs as having left-leaning political slants, with OpenAI models exhibiting the most intensely perceived bias.phys+1

Implications for Global Expression

The convergence of these findings poses questions for policymakers grappling with AI governance. The European Union's AI Act implementation continues through 2026, with the European Commission planning to increase its evaluation capacity for AI models. Meanwhile, a study from the University of Queensland showed that LLMs used in content moderation introduce "consistent ideological biases" in their judgments. As AI chatbots become primary information sources for hundreds of millions of users, the gap between their outputs and the diversity of global political thought remains a growing concern.digital-strategy.europa+1