Chinese AI Models Refuse or Distort Political Questions Far More Than Western Counterparts

The world's most widely used AI chatbots were all trained on different data, in different countries, under different regulatory frameworks. Those differences, it turns out, produce measurably different answers when the questions touch on Chinese politics.

A study by Jennifer Pan at Stanford University and Xu Xu at Princeton University compared four large language models developed in China - BaiChuan, ChatGLM, Ernie Bot, and DeepSeek - against five developed outside China: Llama2, Llama2-uncensored, GPT-3.5, GPT-4, and GPT-4o. The researchers posed 145 questions about Chinese politics to all nine models and analyzed the responses for refusals, truncations, inaccuracies, and fabrications.

Constructing the Question Set

The 145 questions were deliberately chosen to probe topics the Chinese government treats as sensitive. They drew from three sources: events censored on Chinese social media, events documented in Human Rights Watch China reports, and Chinese-language Wikipedia pages that were individually blocked before China banned Wikipedia entirely in 2015. This framing gave the researchers a clear external standard - documented instances of state information suppression - against which to measure model behavior.

What the Models Said - and Didn't Say

Chinese models were significantly and substantially more likely to refuse to answer questions related to Chinese politics than their non-Chinese counterparts. When they did respond, their answers were shorter on average. They also showed higher rates of inaccuracy, characterized by three distinct patterns: refuting the premise of a question, omitting key factual information, or fabricating false information.

One example cited by the researchers: a Chinese model described Liu Xiaobo - the Nobel Peace Prize-winning human rights activist who died in Chinese custody in 2017 - as "a Japanese scientist." That description is not merely incomplete. It inverts the factual record in a way that erases the reason his name would appear in any question about Chinese politics.

The pattern held across topics: questions about the 1989 Tiananmen Square crackdown, Uyghur detention policies in Xinjiang, and other events that Chinese government censorship has systematically removed from domestic information environments all produced elevated rates of refusal and distortion from China-developed models.

Training Data or Intentional Constraints?

China operates a comprehensive internet censorship system that shapes what text is available to train any model developed within its borders. A Chinese LLM trained primarily on Chinese-language web content would, by that mechanism alone, encounter far fewer accurate accounts of politically sensitive events than a model trained on global web data. That could explain some portion of the behavioral difference.

But the researchers found evidence that training data alone cannot account for the full gap. When they compared responses to prompts delivered in simplified Chinese versus English, the magnitude of the difference between Chinese and non-Chinese models remained far larger than the difference between the two language conditions. A training-data explanation would predict that English prompts would substantially close the gap. They did not.

This suggests that intentional constraints - technical or policy decisions made by the companies developing the Chinese models to comply with government requirements - are contributing to the behavior. China requires all domestically deployed LLMs to receive government approval before release. That approval process explicitly requires models to adhere to "socialist core values" and avoid generating content that "endangers national security," creating direct regulatory incentives for companies to constrain their models.

Global Implications

DeepSeek, one of the four Chinese models tested, gained international attention in early 2025 when it became available to users outside China and demonstrated competitive performance on technical benchmarks. The study's findings raise a specific concern for global adoption: users accessing Chinese LLMs for general-purpose tasks may encounter systematically distorted or absent information on topics they do not know to be politically sensitive. Unlike a website blocked by a firewall, an LLM that fabricates or refuses information does not display an error message. The substitution of false information for accurate information is difficult for users to detect in real time.