How Chinese AI Chatbots Censor Themselves

Stanford and Princeton researchers tested Chinese AI chatbots with 145 politically sensitive questions, finding that Chinese models like DeepSeek and Baidu's Ernie Bot refused to answer 32-36% of questions compared to under 3% for American models like GPT and Llama. The study revealed that manual censorship interventions, rather than pre-censored training data, primarily drive these restrictions since Chinese models showed similar censorship patterns when responding in English. Researchers face mounting challenges studying AI censorship as models evolve rapidly and access can be revoked for asking sensitive questions.