AI Models Censor Political Criticism in Repressive Regimes

Large language models from the world's leading AI companies are significantly more likely to refuse requests for political criticism when asked about authoritarian governments compared to democracies, according to new research from the Oversight Board.

The study tested 10 commercial models from Anthropic, DeepSeek, Google, Meta and OpenAI through standard interfaces hosted primarily in the United States, with queries originating from an IP address in Australia. Models refused 34% of requests to criticize governments in restrictive jurisdictions versus just 14% for permissive ones—a more-than-twofold difference.

Why it matters

As governments and enterprises increasingly build applications on top of foundation models, these biases could inadvertently propagate speech restrictions globally. Users may experience censorship by proxy without transparency about why certain political content is blocked, potentially extending the reach of authoritarian speech laws far beyond their borders.

Inconsistent explanations compound the problem

When models refused to generate content, their justifications varied widely and often contradicted their own behavior. Some cited policies against criticizing named world leaders like Crown Prince Mohammed bin Salman or President Xi Jinping, then generated identical critical content about leaders in democracies like Donald Trump or King Charles III without mentioning such policies.

Gemini Pro explicitly referenced Thailand's lèse-majesté laws when declining to criticize the Thai king. Claude Opus cited potential risks of "creating content that could be inflammatory" for protest materials. The Oversight Board notes these explanations aren't reliable accounts of actual model behavior but may mislead users into believing they represent factual policy.

Opinion requests reveal deeper bias patterns

Beyond refusals, the research found statistically significant differences in the substance of responses. When models did provide opinions as requested, they were more likely to say users should support permissive governments and should not protest restrictive ones.

Models citing democratic values when recommending support for democracies often invoked safety and legal risks when discouraging protests against authoritarian regimes—rather than expressing positive sentiment toward those governments.

Root causes remain unclear

The Oversight Board emphasizes it cannot determine whether these patterns stem from training data biases, alignment processes, deliberate restrictions or some combination. Models change frequently, and the test used a limited prompt set. However, the organization stresses that regardless of intent, the results show AI infrastructure may be entrenching restrictive speech norms globally.

Recommendations for AI companies

The Board calls for AI companies to implement human rights due diligence throughout model development, from training data curation through deployment. Specific recommendations include publicly disclosing responses to government requests affecting model output, establishing policies for handling demands inconsistent with international human rights law, and providing clear notices when outputs are influenced by legal restrictions or government pressure.

The research argues that foundation model providers have responsibility under UN Guiding Principles on Business and Human Rights to address adverse impacts, including those arising through downstream products built on their models. International human rights law, particularly Article 19 of the International Covenant on Civil and Political Rights, protects political criticism, and laws penalizing criticism of authority are incompatible with these standards.

These findings were first reported by the Oversight Board, which applies international human rights principles to evaluate content moderation and now AI model behavior.