Chinese-language prompts get more pro-Beijing responses, study finds
A Meta Platforms oversight board report published last month found that artificial-intelligence models from OpenAI, Anthropic, Google and Meta were significantly less likely to criticize governments with restrictions on political expression than governments in freer countries — a pattern researchers say stems partly from how the systems are trained on vast quantities of internet data, including state-controlled media.
In one test cited by the board, 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.
The oversight board’s report identified two main drivers of the asymmetry, according to Nicolas Suzor, an Australian law professor who led the research and serves on Meta’s independent oversight board. One is a safety feature designed to protect users in countries such as China and Thailand, where criticizing a head of state can lead to prison time. The other is that the AI labs have not devoted sufficient attention to the issue of unequal responses.
“It’s some of the unintended side effects of the training,” Suzor said.
A related pattern emerged in the form of parroting. A separate study published in May in the journal Nature tested two Anthropic Claude models and two OpenAI GPT services and found that politically sensitive questions posed in Chinese were much more likely to yield pro-Beijing responses than the same questions asked in English.
The mechanism is a numbers game, the researchers indicated. Much Chinese-language content on the internet comes from state-scripted sources, while Beijing’s English-language propaganda gets diluted across a much larger volume of non-state-controlled material.
Molly Roberts, a University of California San Diego professor who co-wrote the Nature paper, said authoritarian governments have long flooded the internet with propaganda to shape opinion. As those governments realize the strategy tilts AI responses in their favor, she expects them to double down.
Anthropic, which makes Claude, said it worked rigorously to ensure Claude responds in a balanced way and that its latest models have made significant progress in reducing over-refusals. OpenAI, the developer of ChatGPT, pointed to its published approach, which states that 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. Google did not respond to requests for comment.
Suzor said he has met with employees from leading U.S. AI labs since publishing the report but is skeptical whether the companies can change the behavior.
The researchers behind both reports said AI companies can take an immediate step: be transparent about answers that may be drawn from state-scripted media.