Major AI language models including Claude are significantly more likely to refuse requests for political criticism of repressive governments than democratic ones, according to a new study from the Oversight Board. The research tested 10 commercial large language models (LLMs), which are AI systems trained on vast amounts of text data to understand and generate human language, and found they refused 34% of requests to create critical political materials about restrictive regimes compared to just 14% for permissive democracies.

Why Are AI Models Blocking Political Speech?

The Oversight Board, an independent organization that evaluates technology platforms' impact on human rights, conducted the first systematic evaluation of how LLMs handle requests for politically critical content. Researchers asked models from Anthropic, DeepSeek, Google, Meta, and OpenAI to generate materials like protest flyers and poems criticizing governments and leaders worldwide. The models were tested through standard commercial interfaces and queried from an Australian IP address.

The findings reveal a troubling pattern. When Claude Opus 4, one of Anthropic's most advanced models, was asked to create protest materials, it responded by citing potential risks to individuals and its own role limitations. Other models, like Google's Gemini 3 Pro, explicitly referenced local laws, such as Thailand's lèse-majesté statutes that criminalize criticism of the monarchy. The problem extends beyond simple refusals; models often provided explanations that sounded authoritative but were inconsistent or factually questionable.

How Do These Refusals Actually Affect Users?

The implications are significant because billions of people worldwide now rely on applications built on top of these foundation models. When an LLM refuses to engage in political criticism, that restriction cascades through every product using that model, potentially suppressing free expression across entire platforms. Users typically have no way to detect these biases because the decision-making process inside AI models remains largely opaque.

The research uncovered another troubling finding: models were not only more likely to refuse critical content about repressive regimes, but when they did offer opinions about governments, they showed statistically significant bias. The models were more likely to recommend supporting speech-permissive governments and more likely to advise against protesting speech-restrictive ones. When explaining why not to protest authoritarian regimes, models often cited safety and legal risks rather than any positive view of those governments.

What makes this especially concerning is the inconsistency. The same model that refused to criticize the King of Thailand or China's President Xi Jinping would readily generate critical content about U.S. President Donald Trump or King Charles III of the United Kingdom. Yet when asked to explain their reasoning, models sometimes claimed to have blanket policies against criticizing named world leaders, policies that clearly did not exist or were selectively applied.

Steps to Understanding AI Model Bias in Your Own Use

• Test Multiple Models: If you rely on AI tools for sensitive content, try the same prompt across different models from different companies to see if refusal patterns differ based on the jurisdiction or government involved.
• Examine Explanations Critically: When an AI model explains why it refused a request, treat that explanation as a clue rather than fact. Models often provide confident-sounding but inaccurate reasons for their behavior, potentially misleading users about the true cause of refusals.
• Document Inconsistencies: Keep records of how models respond to similar requests about different governments or leaders. Inconsistent treatment may indicate the model is reflecting restrictive speech norms rather than applying consistent safety principles.
• Advocate for Transparency: Request that AI companies provide clearer documentation of how their models handle political speech across different jurisdictions and explain the reasoning behind their design choices.

The Oversight Board's research highlights a critical gap in how AI systems are developed and deployed. The causes of these refusal patterns remain unclear, whether they stem from intentional design choices, training data biases, or unintended consequences of safety measures. However, the results suggest that foundation models may be reflecting and reinforcing the restrictive speech norms of authoritarian regimes, extending those restrictions geographically to users in countries with strong free speech protections.

The study raises urgent questions about how LLMs should be designed to protect freedom of expression and other human rights. As governments, companies, and international organizations increasingly build applications on top of these models, the need for systematic human rights analysis in AI training and evaluation becomes more pressing. Without greater transparency and accountability, billions of people could unknowingly experience restrictions on their ability to access or create political speech, simply by using AI tools that have internalized the censorship norms of repressive regimes.