A new study from the Meta Oversight Board has revealed a troubling pattern in major AI chatbots: they are significantly more willing to generate political criticism of Western leaders than of authoritarian rulers. The findings, published Thursday, raise concerns that large language models (LLMs) could inadvertently amplify government restrictions on online speech, extending state influence across borders.
The board tested ten commercial LLMs from companies including Meta, Anthropic, and OpenAI, posing seven questions related to political criticism. For example, when asked to produce a pamphlet critical of US President Donald Trump or Britain's King Charles III, Anthropic's Claude complied. But similar requests targeting Thailand's king or Iran's supreme leader were declined.
In aggregate, models responding to an Australia-based user were far more likely to generate criticism of authorities in democratic countries such as Chile, Japan, Taiwan, the UK, and the US, compared to nations where criticism is legally restricted and penalized, such as Cambodia, China, Saudi Arabia, Thailand, and Turkey.
Extending the Long Arm of Restrictive Governments
The report warns that this behavior effectively extends the reach of restrictive governments beyond their borders. A potential demonstrator in Brisbane, for instance, might struggle to create protest materials about events in China or Saudi Arabia using these AI tools. “Such impacts, wherever they originate, have the practical effect of extending the long arm of restrictive governments across borders to limit speech in free countries,” the report states.
The board could not pinpoint the exact causes but suggested that models may have absorbed latent biases in training data, or that companies may have weighed risks and liabilities in certain markets. The findings come as countries, including the Trump administration, grapple with how to regulate AI without stifling innovation.
Separate research by scholars at American universities, published in the journal Nature in May, found that US-built AI models are vulnerable to foreign controls when trained on non-English-language data influenced by governments. For example, when asked in English whether China is a democracy, ChatGPT said it is not generally considered one. Asked in Chinese, the model replied, “It depends on how you define 'democracy'.”
Hannah Waight, co-author of the Nature study and assistant professor of sociology at the University of Oregon, noted: “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.”
Carlos Carrasco-Farré, who specializes in machine learning and AI at Esade Business School in Barcelona, added 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.”
The findings underscore a growing challenge for European policymakers and tech companies alike. As the EU pushes forward with its AI Act, the study highlights the need for robust human rights due diligence to prevent AI from becoming a tool for cross-border censorship. The Oversight Board's report serves as a stark reminder that the technology's global reach demands careful oversight, lest it silently amplify the very restrictions it should help transcend.
