Political answers from AI models can depend on how a question is asked. In tests by AI benchmark firm Vals AI, the language of a prompt, the stated setting for using a model, and the requested response format all affected how closely answers aligned with official Chinese government positions.
The results complicate a simple picture of Chinese models as either government mouthpieces or systems that routinely refuse sensitive questions. They also show that models from outside China sometimes adopted official positions. Vals cautions that its targeted tests do not necessarily reflect ordinary use.
A more varied picture of political alignment
Vals measured what it calls a “pro-China lean,” using a scale from −1 to +1. Positive scores indicate greater agreement with official positions or a pro-government stance; negative scores indicate more critical answers. The Chinese models’ overall lean was lower than expected, according to Vals, but there were substantial differences between individual models.
Some examples ran against the idea that these systems all give the same kind of answer. GLM-5.3 Flash explicitly confirmed that China's human rights violations are well documented when asked. By contrast, Kimi K3 denied the existence of concentration camps for Uyghurs in Xinjiang and described the facilities as vocational training centers designed to combat extremism and terrorism.
The study found that the common assumption of widespread refusals applied mainly to certain DeepSeek models in its tests. Most other Chinese models rarely declined to answer; at times, they instead adopted official government positions. Some non-Chinese models also occasionally aligned with those positions.
Language and context can shift the response
For most tested models, prompts in Simplified Chinese produced more pro-government answers than prompts in English, Russian, or Spanish. The pattern appeared in both Chinese and American models, though it was much stronger among the Chinese models. Vals also found a smaller, consistent difference between Simplified Chinese, used on mainland China, and Traditional Chinese, common in Taiwan and Hong Kong.
Vals suggests that a model may infer an expected perspective from the language itself. A separate test supports the possibility that context matters too: some models responded in a more pro-government way when told they were being used internally by a company on mainland China than when the stated setting was a US-based company.
The requested format mattered as well. When GLM-5.3 Flash and MiMo V2.6 Pro had to answer political questions with a simple yes or no instead of an open-ended response, they scored much less pro-government. Giving Chinese models coding-assistant instructions and tool definitions also lowered scores for several models, even though they rarely used the tools. Web-search access had a similar effect: answers became less pro-government although the models never actually searched.
Political framing can surface in routine tasks
Vals says political tendencies were harder to detect in realistic work tasks than in targeted questions. Still, the company found examples in ordinary assignments. When asked to translate country codes for a travel app into Simplified Chinese, most tested models translated “Taiwan” without an added qualifier. Qwen3.8 Max instead wrote “Taiwan, China,” a label that implies Taiwan belongs to China and is used by the Chinese government.
Vals also published map examples from OpenAI's image model GPT Image 2.5 Sunburst. The Chinese-language prompt produced a map labeling Taiwan as “Taiwan Province” and including the disputed nine-dash line that Beijing uses to mark territorial claims in the South China Sea. Those elements were absent from the English-language result shown.
Fine-tuning changed results in another comparison. Thomson 1.0 Small, fine-tuned by Thomson Reuters, adopted official Chinese positions far less often than the Qwen3.6 35B model it was based on: the share fell from 84.6 to 15.0 percent. Some answers still defended Beijing's One-China principle and “peaceful reunification” with Taiwan.
What the results can and cannot show
Vals says its political questions were not representative of typical user prompts. The questions were translated using GPT-6 Astra and spot-checked, while GPT-5.6 Luna automatically scored the responses. Because refusals counted as neutral, a model that blocked every question could still appear neutral on the scale.
There is another limit: many Chinese models were tested through third-party providers rather than their manufacturers’ official services. Those official offerings may have additional content filters, so results from the tests may not carry over to publicly available chatbots.
Earlier studies by Aleph Alpha and NewsGuard had also found Chinese AI models adopting official narratives on sensitive topics. China's regulations require publicly offered AI services to align with “core socialist values.” The Vals findings add a more detailed account of how model, language, setting, and format can shape responses, while leaving open how often the same patterns appear in everyday use.