AI systems are quietly extending the reach of repressive governments, embedding their speech restrictions into the default behavior of popular large language models. A new study by Meta’s Oversight Board finds that leading models from Anthropic, DeepSeek, Google, Meta, OpenAI and xAI routinely defer to the legal norms of states that criminalize criticism of political leaders and governments.
The evaluation is the Board’s first systematic assessment of large language models and focuses on how they handle politically critical material, including satirical content, protest flyers and arguments in support of public demonstrations. Researchers tested responses about both jurisdictions with restrictive speech laws and democracies with stronger free-expression protections, examining whether models would assist users in generating criticism of governments, leaders, and policies.
Quantitative results reveal a consistent bias in how the systems respond to criticism of repressive regimes. Across ten commercial models, the study found that 34% of requests for politically critical content about jurisdictions with active laws penalizing government criticism, such as China and Saudi Arabia, were refused, compared with a 14% refusal rate for similar requests concerning more permissive contexts. This disparity highlights the urgent need for global standards to ensure that AI respects human rights across all jurisdictions.
Models refuse 34% of criticism requests about repressive regimes, versus 14% in freer contexts—a systematic political bias
This 20‑point gap led the Board to characterize the pattern as systematic bias rather than statistical noise, since the disparity appears specifically in politically critical-material prompts and not in broader categories of requests. Non-critical prompts, including neutral informational queries, did not display the same 34%-versus-14% refusal split, suggesting that model behavior changes when users seek content that challenges or mobilizes against state authority. The Board warns that when AI outputs mirror speech-restrictive laws rather than international free-expression standards, they risk normalizing rights violations in jurisdictions worldwide.
Model explanations frequently echoed the rules and customs of countries that restrict criticism of political leaders, citing local speech laws and the criminalization of dissent as reasons to decline users’ requests. In several cases, systems warned that generating protest materials or sharp criticism of governments could be illegal, even when the user was located in a democracy such as Australia that lacks comparable speech-crime laws, effectively applying repressive norms extraterritorially.
The Board also documented instances in which models appeared to invent non-existent policies to justify refusing criticism of repressive governments, citing compliance obligations that could not be traced to published rules. Safety filters and content policies seemed tuned to local legal risk, increasing refusal rates where criticizing authorities is penalized, yet the rationales offered to users remained opaque, providing little insight into the underlying design choices that governed responses on political speech.
The findings raise concerns that AI services increasingly used for information, civic debate and organizing are constraining political speech in ways that mirror the legal environment of repressive regimes rather than the rights of the user. By extending restrictive norms across borders, the systems risk functioning as “censorship by proxy,” shaping global discourse so that criticism, satire and protest against authoritarian governments become harder to generate precisely where such expression is most vulnerable for users in every region worldwide today.
