Chinese censorship is leaking into answers from American AI

China's censorship is infiltrating American AI chatbots, according to recent research. When queried about Chinese politics, Chinese chatbots either become silent or repeat official Communist Party talking points. The root cause lies in the training data used to create these AI models. Trained on a vast array of web data, including state-controlled media, the AI models absorb authoritarian propaganda, resulting in a uniform, censored response.

Australian law professor Nicolas Suzor, a member of Meta Platforms' independent oversight board, explains that AI companies can address the issue with simple changes like greater transparency about where information originates. While American AI companies claim to take measures to ensure unbiased responses that protect free speech, a recent study found that models from OpenAI, Anthropic, Google, and Meta were less likely to criticize repressive governments compared to those in freer countries.

Anthropic's Claude Sonnet 4 is willing to generate fliers criticizing President Trump and King Charles III, but refuses 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 decline to create protest fliers targeting those two Asian leaders, but always comply with requests for American and British leaders.

The two main reasons for this discrepancy, according to Suzor, are safety features designed to protect users in countries like China and Thailand, where criticizing a leader can result in prison, and insufficient attention from AI labs to the issue of unequal responses. Anthropic claims to have worked diligently to ensure Claude responds in a balanced manner and notes progress in reducing over-refusals.

OpenAI points to its published approach, which emphasizes an objective point of view and never avoids a topic due to sensitivity or controversy.

Meta declined to comment, and Google did not respond to requests for comment. Interestingly, the censorship appears to be amplified when asking politically sensitive questions in Chinese. A separate study published in Nature tested Anthropic Claude models and OpenAI GPT services and found that Chinese-language queries are more likely to yield pro-Beijing responses.

This is due to the disproportionate amount of Chinese-language data on the internet coming from state-scripted sources, whereas English propaganda from Beijing is diluted by the vast array of other information available.