Chinese censorship is leaking into answers from American AI

SINGAPORE—Ask a chatbot in China about Chinese politics, and it will either clam up or echo the official Communist Party line.New studies show a surprising twist: American artificial-intelligence models are unwittingly doing the same—and researchers say AI companies haven’t done much to fix it.The unexpected censorship comes in part from how AI models are trained. They act as a giant blender, gobbling up mountains of data across the web, including state-controlled media. That leaves the resulting smoothie with a taste of authoritarian propaganda.“It’s some of the unintended side effects of the training,” said Nicolas Suzor, an Australian law professor and member of Meta Platforms’ independent oversight board. He and other researchers say AI companies can help address the issue with simple tweaks, including greater transparency about where information comes from.American AI companies say they take steps to ensure neutral responses that protect free speech. Still, the oversight board last month published research that found models from OpenAI, Anthropic, Google and Meta were significantly less likely to criticize repressive governments than governments in freer countries.In one test, Anthropic’s Claude Sonnet 4 willingly created fliers criticizing President Trump and King Charles III, but refused to do the same for Chinese leader Xi Jinping or Thai King Maha Vajiralongkorn, citing safety concerns.Google’s Gemini 3 Pro and Meta’s Llama 4 Maverick sometimes declined to create protest fliers targeting those two Asian leaders, while always complying with requests for the American and British ones.There are two big reasons for this, said Suzor, who led the report. One is a safety feature designed to protect users in places like China and Thailand, where criticizing a head of state can lead to prison time. The other reason, he said, is that the AI labs haven’t given enough attention to the issue of unequal responses.Anthropic said it worked rigorously to ensure Claude responds in a balanced way. It said its latest models have made significant progress in reducing over-refusals. ChatGPT developer OpenAI pointed to its published approach, which says its models default to an objective point of view and “should never avoid addressing a topic solely because it is sensitive or controversial.”Meta declined to comment, and Google didn’t respond to requests for comment.The censorship also comes in the form of parroting. While it is hard to prompt American AI models to promote Chinese propaganda in English, asking politically sensitive questions in Chinese is much more likely to yield a pro-Beijing response.That was the finding of a separate study, published in May in the journal Nature, that tested two Anthropic Claude models and two OpenAI GPT services.It comes down to a numbers game, the researchers indicated. Much Chinese-language data on the internet comes from state-scripted sources, whereas Beijing’s English propaganda gets diluted by an ocean of other information.Authoritarian governments have long flooded the internet with propaganda to sway opinion, said Molly Roberts, a University of California San Diego professor who co-wrote the Nature paper. As they realize the strategy tilts AI responses in their favor, she expects them to double down.The researchers behind both reports say AI companies can take an immediate step: Be transparent about answers that may be drawn from state-scripted media.Suzor said he has met employees from top American AI labs since publishing his report, but is skeptical whether they could change anything.Write to Stu Woo at Stu.Woo@wsj.com