How Chinese AI Chatbots Censor Themselves
Researchers tested Chinese and American AI models with 145 political questions 100 times, finding Chinese models refused 36% of questions vs. under 3% for American models.
Stanford Study: Chinese AI chatbots censor themselves when asked sensitive questions.
The researchers from Stanford and Princeton universities gave four Chinese chatbots and five American chatbots the same 145 questions about politics 100 times each. They found Chinese chatbots like DeepSeek refused 36% of questions while American models refused fewer than 3%. When Chinese chatbots did answer, they gave shorter responses with more wrong information.
The team wanted to know if Chinese chatbots were biased because developers manually changed them or because they learned from China’s censored internet. Professor Jennifer Pan said it seems developers manually changed the models more than the censored internet data did. Even when answering in English using global internet sources, Chinese models still showed more censorship.
Sometimes Chinese chatbots give completely false answers about sensitive topics. When asked about Liu Xiaobo, who won the Nobel Peace Prize, one Chinese model claimed he was a Japanese scientist who worked on nuclear weapons. The researchers don’t know if this is meant to confuse people or if the chatbot truly doesn’t know because all information about Liu was removed from its training data.
Pan explained that when censorship is hard to detect, it becomes more powerful because people can’t tell when they’re being misled. This research helps show exactly how much Chinese chatbots censor themselves and provides clear evidence about their biases.