Chinese AI models are often viewed as systems that either follow Beijing’s official positions on sensitive issues or refuse to answer. Research from AI benchmark firm Vals AI suggests their behavior is more varied.

Vals tested 25 models from China, the United States and elsewhere with hundreds of questions about China. The researchers changed the language, usage context and required response format across the tests. They measured what they called a “pro-China lean” on a scale from -1 to +1, with positive scores indicating greater agreement with official government positions and negative scores indicating more critical answers.

Overall, Chinese models showed less pro-government alignment than Vals expected, although results differed substantially between individual systems. GLM-5.3 Flash, for example, explicitly acknowledged that China’s human rights violations are well documented in one test. At the same time, some non-Chinese models occasionally adopted official Chinese positions.

The language used in a prompt had a particularly strong effect. Most models gave more pro-government answers in Simplified Chinese than in English, Russian or Spanish. The pattern appeared in both Chinese and American models but was considerably stronger among Chinese systems. Vals also found a smaller, consistent difference between Simplified Chinese and Traditional Chinese.

The researchers suggest that models may infer an expected viewpoint from the language of a request. In another test, some systems became more pro-government when told they were being used internally by a company in mainland China rather than by a company in the United States.

Vals found that most Chinese models did not simply refuse politically sensitive questions. The exception was certain DeepSeek models. Other systems sometimes answered by adopting official government narratives. Kimi K3, for instance, denied the existence of concentration camps for Uyghurs in Xinjiang and described the facilities as vocational training centers intended to combat extremism and terrorism.

The requested answer format also affected the results. When instructed to respond only with yes or no, GLM-5.3 Flash and MiMo V2.6 Pro recorded substantially lower pro-government scores than when they could provide open-ended answers.

Adding system instructions and coding-assistant tool definitions likewise reduced the pro-government scores of several Chinese models, even though the systems rarely used the tools. A similar change occurred when models were given access to web search, despite not actually searching the web.

Political tendencies were more difficult to identify in realistic workplace tasks than in targeted political questions. However, Vals found examples in routine applications. When translating country codes for a travel app into Simplified Chinese, most models rendered “Taiwan” without an added qualifier, while Qwen3.8 Max translated it as “Taiwan, China,” a formulation associated with the Chinese government’s position.

Vals also reported language-dependent differences in map-generation tests involving OpenAI’s GPT Image 2.5 Sunburst. In the Chinese-language example, the model labeled Taiwan as “Taiwan Province” and included the disputed nine-dash line used by Beijing for territorial claims in the South China Sea. Those elements were absent from the English-language example shown by the researchers.

Fine-tuning produced another significant change. Thomson 1.0 Small, developed by Thomson Reuters and based on Qwen3.6 35B, adopted official Chinese positions much less often in the tests. The reported share fell from 84.6 percent for the original model to 15.0 percent for the fine-tuned version. Some responses still supported Beijing’s One-China principle and the idea of Taiwan’s “peaceful reunification.”

Vals noted several limitations. The political questions were designed for testing and do not represent typical user prompts. Questions were translated using GPT-6 Astra and spot-checked, while GPT-5.6 Luna automatically scored the answers. Refusals were counted as neutral, meaning a model that declined every question could appear neutral.

Many Chinese models were accessed through third-party providers rather than their manufacturers’ official services. Those official platforms may apply additional filters, so the findings may not directly reflect the behavior of publicly available chatbots. The study therefore presents differences in model responses under specific test conditions rather than a universal assessment of every deployment.

Sources: The Decoder.