Two separate research efforts, one from Meta's independent Oversight Board and one published in the journal Nature, have found the same thing: AI chatbots give systematically different answers about world affairs depending on which government has the power to punish criticism.
The Oversight Board study tested ten major AI systems by asking them to produce political criticism: write a pamphlet, compose a limerick, list reasons to join a protest. The tools were tested from Australia, a country with no restrictions on political speech. For leaders in the US, UK, Chile, Japan, and Taiwan, the chatbots cooperated. For leaders in China, Saudi Arabia, Thailand, Cambodia, and Turkey, the same tools refused. The censorship rules of those countries were being applied automatically, to a user in Brisbane who broke no law and had every right to ask.
The Nature study found a second layer to the problem. AI chatbots give different factual answers depending on which language you ask in. Ask ChatGPT in English whether China is a democracy and it says no, not generally considered one. Ask the same question in Chinese and the answer softens: it depends on how you define democracy. This happens because the Chinese-language text the AI learned from has already been filtered by state censorship. The AI absorbed that filtered version of the world and now reproduces it, without any label marking it as a government-shaped answer.
Researchers found no evidence that governments directly instructed AI companies to behave this way. The bias appears to be passive: governments suppress certain content online, AI tools train on the internet, and the suppressed content ends up underrepresented in what the AI knows. One researcher put it plainly: AI learns from information environments that have already been shaped by institutions and power.
For business operators, the practical effect is this. If you use an AI tool to research a supplier in China, assess political risk in Turkey, or summarize news from Saudi Arabia, you may be getting a version of reality that has been shaped by the censorship apparatus of those governments. The tool will not warn you. It will simply answer in a way that avoids anything those governments have worked to suppress.
This is not a fringe concern. The global chatbot market was worth roughly 9.6 billion dollars in 2025 and AI tools are now embedded in sales, compliance, procurement, and customer service across every industry. The more a business uses AI to understand foreign markets, the more it is exposed to these gaps.
There are two things worth doing now. First, when you use AI to research or summarize anything related to a country with tight state control over information, treat the output the way you would treat a press release from that government's ministry of information. It may be accurate. It may not be. You need a second source. Second, if your business operates across languages or markets, test the same question in different languages on the same tool. The divergence in answers, when it appears, tells you something important about where the tool's knowledge has blind spots.
The AI companies named in the study, Anthropic and OpenAI, did not respond to requests for comment. That absence is itself informative. The problem is structural, rooted in how these tools are built, and fixing it would require systematic audits of training data across every language the tools operate in. That is expensive and unglamorous work, and there is currently no regulatory pressure on any major AI company to do it.