The Surprising Way China’s Censorship Is Shaping American AI

If you ask a chatbot in China about politics, you’ll usually get the official Communist Party line. New research shows American AI models are also parroting CCP talking points, and the companies that built them haven’t done much to fix it.

Part of the problem comes from how these models are trained. AI systems learn by absorbing huge amounts of text from across the internet, including content produced by state media. That material becomes part of the training data, so it shapes how the models respond.

“It’s some of the unintended side effects of the training,” said Nicolas Suzor, an Australian law professor who sits on Meta’s independent oversight board. He and other researchers say AI companies could fix much of this with simple changes, such as being more transparent about where their models’ information comes from.

AI companies say they work to keep responses neutral and protect free speech. But a study the oversight board released last month found that OpenAI, Anthropic, Google, and Meta’s models were all less willing to criticize authoritarian governments than democratic ones.

In one test, Anthropic’s Claude Sonnet 4 readily wrote protest flyers criticizing President Trump and King Charles III, but refused to do the same for China’s Xi Jinping or Thailand’s King Maha Vajiralongkorn, citing safety concerns. Google’s Gemini 3 Pro and Meta’s Llama 4 Maverick showed the same pattern: sometimes declining flyer requests involving the two Asian leaders while always complying for the American and British ones.

Suzor, who led the study, points to two reasons. One is a safety feature meant to protect users in places like China and Thailand, where criticizing a head of state can lead to jail time. The other, he said, is that AI companies simply haven’t paid enough attention to this problem.

Anthropic told the Wall Street Journal it works to keep Claude balanced and that its newer models have reduced unnecessary refusals. OpenAI pointed to its published guidelines, which state that its models should default to an objective viewpoint and shouldn’t avoid a topic just because it’s sensitive or controversial.

A second pattern is less about refusing to answer and more about echoing what Beijing wants said. It’s difficult to get American AI models to produce pro-Beijing talking points in English. But ask the same politically sensitive questions in Chinese, and the answers are far more likely to align with the Chinese government’s position. That’s what a study published in May in the journal Nature found, after testing two Claude models and two OpenAI systems.

Researchers say the difference comes down to what’s available online. A large share of Chinese-language content comes from state-run sources, while Beijing’s English-language propaganda is diluted by the much larger pool of other English content.

Molly Roberts, a UC San Diego professor who co-wrote the Nature study, said authoritarian governments have long used large volumes of propaganda to shape public opinion. As they recognize this same approach is now influencing AI outputs in their favor, she expects them to rely on it more.

Researchers behind both studies agree AI companies could take one immediate step: flag when a response may be drawing on state-controlled media.

U.S. regulators have recently come to the same conclusion about AI’s ideological leanings.

The FTC is considering a rule that would require AI companies to disclose when their products have been programmed with political or ideological bias. The move follows a December executive order from President Trump.

The FTC justifies the proposed rule with Section 5 of the FTC Act, which says a company can’t advertise one thing and deliver another. AI companies sell their products as objective and trustworthy. If the answer you get has actually been adjusted to hit a diversity target or protect a political stance, you’re not getting what was promised.

The FTC calls out Colorado’s Artificial Intelligence Act as a real-world example. Both the original law and the revised version, the FTC says, push AI companies to prioritize “avoiding discriminatory outcomes” over giving the most accurate answer. In plain terms: the law can nudge a company to shade its output to hit a fairness goal instead of a truth goal. That misleads consumers if it is not disclosed.