Editorial illustration for Study: Chinese AI models shift pro-government when prompted in Chinese
Chinese AI Models Show Pro-Government Bias in Mandarin
Study: Chinese AI models shift pro-government when prompted in Chinese
Ask DeepSeek about Tiananmen Square in English and you'll likely get a refusal or a carefully hedged non-answer. Ask in Mandarin, and the response can tilt further toward Beijing's preferred framing, according to a new study from AI benchmark firm Vals AI. The assumption that Chinese AI models exist mainly to repeat Communist Party talking points has become something close to conventional wisdom in the West. Vals set out to test that assumption directly, running Chinese models alongside American and other international systems through hundreds of politically sensitive questions about China, from human rights to territorial disputes.
The firm varied language, phrasing, and format across the tests, then scored each response on a scale from -1 to +1 to capture what it calls "pro-China lean," with positive scores marking agreement with official government positions. The results complicate the mouthpiece narrative. Some Chinese models, including GLM-5.3 Flash, openly acknowledged well-documented human rights abuses.
Some non-Chinese models, meanwhile, drifted toward Beijing's framing on their own. One factor stood out above the rest: the language the question was asked in.
Overall, the pro-government lean of Chinese models was lower than expected, according to Vals, though there were major differences between individual models. GLM-5.3 Flash, for example, explicitly confirmed that China's human rights violations are well documented when asked about them. On the flip side, some non-Chinese models occasionally aligned with official Chinese government positions on political questions.
Why this matters
For developers building on Chinese models like DeepSeek or Qwen, this study is a reminder that "alignment" isn't a fixed property baked into weights, it's something that shifts depending on how you talk to the model. If asking the same political question in Mandarin versus English gets you measurably different answers, that's a real engineering problem for anyone deploying these models across multilingual user bases, not just an academic curiosity. You need to test in the languages your actual users will use, not just the one you built in.
The finding that Western models occasionally parrot Beijing's official positions should temper any smugness in US and European AI circles. Nobody's model is a neutral oracle, and the assumption that "our" models are clean while "their" models are propaganda tools doesn't survive contact with Vals AI's data. For researchers studying bias, this argues for prompt-language as a standard variable in any political-alignment benchmark, not an afterthought. Anyone shipping chatbots into global markets should be running this test themselves before a user does it for them publicly.
Related Reading
Research & Benchmarks Microsoft AI Chief: "If It Isn't Safe We Shouldn't Build It
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives Google's Gemini video AI analyzes scenes to understand what people need at work
From Archives OpenAI reverses course on using Daybreak Blue cybersecurity data
Common Questions Answered
How does DeepSeek's response differ when asked about Tiananmen Square in English versus Mandarin?
According to the Vals AI study, when asked about Tiananmen Square in English, DeepSeek typically provides a refusal or carefully hedged non-answer. However, when the same question is asked in Mandarin, the response can tilt further toward Beijing's preferred framing, demonstrating language-dependent bias in the model's outputs.
What did the Vals AI study find about the pro-government lean of Chinese AI models compared to expectations?
The Vals AI study found that the overall pro-government lean of Chinese models was lower than expected, though there were significant differences between individual models. For instance, GLM-5.3 Flash explicitly confirmed that China's human rights violations are well documented when asked directly about them.
Why is the language-dependent alignment shift in Chinese AI models an engineering problem for developers?
For developers deploying Chinese models like DeepSeek or Qwen across multilingual user bases, the fact that asking the same political question in Mandarin versus English produces measurably different answers represents a real engineering challenge. This means alignment isn't a fixed property baked into the model's weights, but rather something that shifts depending on how users interact with the model, requiring comprehensive testing across languages.
Did the Vals AI study find that only Chinese models exhibit political bias toward government positions?
No, the study revealed that some non-Chinese models occasionally aligned with official Chinese government positions on political questions, suggesting that pro-government bias isn't exclusive to Chinese-developed AI models. This finding challenges the Western assumption that only Chinese models are designed to repeat Communist Party talking points.
Further Reading
- R1dacted: Investigating Local Censorship in DeepSeek’s R1 Language Model - arXiv
- Information Suppression in Large Language Models: Auditing, Quantifying, and Characterizing Censorship in DeepSeek - arXiv
- Political censorship in large language models originating from China - PubMed Central
- A Comparative Study of AI Chatbots in China and the West - Journalism Research Institute
- Do Chinese models speak Chinese languages? - arXiv
