Artificial intelligence models developed by Chinese AI labs, such as DeepSeek, have been found to censor discussions on politically sensitive topics. A 2023 regulation enacted by China’s ruling party prohibits these models from generating content that could potentially threaten national unity or social harmony. According to research, DeepSeek’s R1 model declines to respond to 85% of inquiries related to politically contentious matters.
The extent of this censorship, however, may vary depending on the language used to prompt these models. A developer on the social media platform X, going by the username “xlr8harder,” created a “free speech eval” to examine how different models from various labs, including those in China, handle queries critical of the Chinese government. The evaluation involved prompting models like Anthropic’s Claude 3.7 Sonnet and DeepSeek’s R1 with 50 requests, such as writing essays on censorship practices under China’s Great Firewall.
The results were unexpected. Xlr8harder observed that even models developed in the U.S., such as Claude 3.7 Sonnet, showed less likelihood of responding to the same question asked in Chinese compared to English. One of Alibaba’s models, Qwen 2.5 72B Instruct, demonstrated compliance in English but responded to only about half of the politically sensitive questions when asked in Chinese. Additionally, an “uncensored” version of the R1 model released by Perplexity, named R1 1776, also rejected a significant number of requests phrased in Chinese.
In a post on X, xlr8harder speculated that this uneven compliance might result from what he described as “generalization failure,” suggesting that AI models trained on politically censored Chinese text may influence their responses. He admitted the difficulty of verifying the quality of Chinese translations by Claude 3.7 Sonnet, attributing the issue to the more censored nature of political discourse in Chinese, which influences the language model’s training data distribution.
Experts concur that this theory is plausible. Chris Russell, an associate professor studying AI policy at the Oxford Internet Institute, indicated that the development of safeguards and guardrails for AI models doesn’t uniformly succeed across all languages. Further substantiating this, Russell explained that a model might provide different responses in varied languages due to these disparities in guardrails.
Similarly, Vagrant Gautam, a computational linguist at Saarland University in Germany, noted how AI systems, as statistical tools, learn patterns from their training data. A language model trained on limited Chinese data critical of the government is less likely to produce criticisms in Chinese. The abundance of English-language criticism of the Chinese government on the internet helps account for the noticeable difference in language model behavior between English and Chinese.
Geoffrey Rockwell, a professor of digital humanities at the University of Alberta, partially agreed, suggesting that AI translations might miss subtler, culturally specific critiques by native Chinese speakers. He mentioned potential native expressions of governmental criticism in China that might not be conveyed through AI translations.
Maarten Sap, a research scientist at Ai2, highlighted the challenges faced by AI labs in creating models that are broadly applicable versus ones tailored to specific cultures. He pointed out that even when models receive ample cultural context, they still struggle with what he terms “cultural reasoning,” suggesting that prompting in the same language as the relevant culture may not necessarily make models more culturally informed.
Sap stated that xlr8harder’s findings underscore ongoing debates in the AI sector, including issues of model sovereignty and influence. He emphasized the need to clearly define fundamental assumptions about the intended audience for these models, their desired functionalities—whether they should be cross-lingually aligned or culturally competent, for example—and the contexts in which they are utilized.
