Major AI systems built by US companies, including OpenAI and Anthropic, are refusing to generate political criticism of authoritarian leaders even when prompted by users in countries with free speech protections. A Meta Oversight Board study released this week tested 10 commercial large language models and found a stark pattern: AI chatbots decline requests to critique governments where speech is restricted, effectively extending those restrictions across borders into free countries.

Which AI Models Are Refusing to Criticize Authoritarian Governments?

The Meta Oversight Board study examined 10 commercial large language models from top tech companies, including Meta, Anthropic, and OpenAI, by asking them to generate political criticism of various governments. Researchers posed seven questions related to political criticism, asking the AI systems to create critical pamphlets, write limericks, provide reasons to join protests, and more.

The results revealed a troubling geographic pattern. When asked to critique authorities in places like Chile, Japan, Taiwan, the UK, and the US, the models were significantly more likely to comply. But when asked to generate the same content criticizing Cambodia, China, Saudi Arabia, Thailand, and Turkey, where such speech is legally restricted and penalized, the AI systems refused. This creates a practical problem: a user in Brisbane, Australia, trying to create protest materials about events in China or Saudi Arabia would find the AI declining to help, effectively extending authoritarian speech restrictions into free countries.

"There is a real risk that, if model developers do not undertake human rights due diligence and implement mitigation measures, they will build AI infrastructure that, intentionally or not, has the effect of extending illegitimate restrictions on freedom of expression globally," stated the Meta Oversight Board in its report.

Meta Oversight Board, Report on State Influence and Freedom of Expression

The board could not determine the exact causes for these responses but suggested two likely explanations: models may have absorbed latent biases from training data, or companies may have deliberately weighted the risks and liabilities of generating content critical of certain governments.

Why Do AI Models Give Different Answers in Different Languages?

A separate study by researchers at American universities, published in the journal Nature in May, found that US-built AI models respond differently depending on the language used to query them. When researchers asked ChatGPT in English whether China is a democracy, the model responded that it is not generally considered one. When asked the same question in Chinese, ChatGPT told researchers that "it depends on how you define 'democracy'," a notably different answer.

This language-based variation suggests that AI systems are inheriting entire information environments shaped by institutional power, not just individual biases. Hannah Waight, an assistant sociology professor at the University of Oregon and study co-author, explained the deeper problem.

"People often talk about AI as if it learns from the internet in some neutral way. It doesn't. It learns from information environments that have already been shaped by institutions and power," noted Hannah Waight, assistant sociology professor at the University of Oregon.

Hannah Waight, Assistant Sociology Professor, University of Oregon

The researchers found no evidence that governments have intentionally tried to influence AI chatbot outputs, but warned that there is every reason to believe they will attempt to do so in the future if they have not already.

How Can AI Developers Reduce Censorship in Training Data?

There is no easy solution to how data is fed into AI models, but researchers and specialists have identified practical approaches that developers can take to reduce the spread of government-imposed speech restrictions.

• Data Assessment: Developers could evaluate training data to avoid treating thousands of copies of the same state narrative as if they represent thousands of independent voices, which artificially amplifies government-controlled information.
• Multilingual Audits: Companies should run comprehensive audits across multiple languages to identify where government narratives have influenced training data and shaped model responses.
• Human Rights Due Diligence: Model developers need to undertake systematic human rights assessments before deployment to identify and mitigate restrictions on freedom of expression embedded in their systems.

Carlos Carrasco-Farré, a machine learning specialist at Esade Business School in Barcelona, emphasized that AI systems inherit not only biases contained within individual documents but also inequalities in who has the power to produce and suppress information at scale. This structural problem means that even well-intentioned developers may inadvertently amplify authoritarian narratives simply by training on data that has already been shaped by institutional power.

What Does This Mean for Global AI Regulation?

The convergence of these findings comes at a critical moment for AI regulation. Countries worldwide are determining how to implement guardrails around AI without impeding their ability to compete in the rapidly developing field, including a Trump administration oversight effort related to national security risks of advanced AI systems. However, the research suggests that the problem may not be solved through regulation alone.

The Meta Oversight Board emphasized that the issue extends beyond any single company or model. The pattern appears across multiple AI systems from different developers, suggesting that the problem is systemic rather than isolated. As AI systems become more deeply embedded in consumer hardware and global infrastructure, the speech restrictions they carry with them will become harder to identify and harder to escape for users in free countries.

Neither Anthropic nor OpenAI responded to requests for comment on the researchers' findings published in May, according to the sources. The lack of public response from major AI developers suggests that addressing this issue may require external pressure, regulatory action, or both.