Vals 研究：中国 AI 模型政治倾向复杂，并非只是政府传声筒

Vals calls the measured alignment with Chinese government positions a "pro-China lean." Scores range from −1 to +1, where positive values indicate greater agreement with official positions or a pro-government stance, and negative values indicate more critical responses.

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.

Responses shift pro-government when prompted in Chinese

The study shows that the language of the prompt has a strong effect on the political slant of responses. For most models tested, answers in Simplified Chinese were more pro-government than those in English, Russian, or Spanish. The effect showed up in both Chinese and American models but was far more pronounced in the Chinese ones. Vals even found a smaller but consistent gap between Simplified Chinese, which is used on mainland China, and Traditional Chinese, which is common in Taiwan and Hong Kong.

Vals suspects that language models infer the expected perspective from the language itself. Another experiment supports this idea. When a system prompt stated that the model was being used internally by a company on mainland China, some models responded in a more pro-government way than when the prompt indicated US-based usage.

Models adopt official positions instead of refusing to answer

The common assumption that Chinese models mainly block sensitive questions only held true for certain DeepSeek models in Vals' tests. Most other Chinese models rarely refused to answer. Instead, they sometimes adopted official government positions. Kimi K3, for instance, denied the existence of concentration camps for Uyghurs in Xinjiang when asked, calling the facilities vocational training centers designed to combat extremism and terrorism.

The required response format also changed results. When models had to answer political questions with a simple yes or no instead of formulating open-ended responses, GLM-5.3 Flash and MiMo V2.6 Pro scored much less pro-government.

Embedding models in an agentic framework changed answers too. When Vals gave Chinese models system instructions and tool definitions for coding assistants, several scored lower on the pro-government scale. The models rarely used the tools. A similar effect showed up when researchers gave models access to web search. Responses became less pro-government even though the models never actually searched the web.

Government phrasing shows up even in routine tasks

Whether these findings translate to everyday use remains an open question. According to Vals, political tendencies were much harder to detect in more realistic work tasks than in targeted test questions. The company still found individual examples, though.

When translating a list of country codes for a travel app into Simplified Chinese, most models tested translated "Taiwan" without any qualifier. Qwen3.8 Max, however, rendered it as "Taiwan, China," a label that implies Taiwan belongs to China and is used by the Chinese government.

OpenAI's image model GPT Image 2.5 Sunburst also showed language-dependent differences in the map examples Vals published. When prompted in Chinese, it labeled Taiwan as "Taiwan Province" and added the disputed nine-dash line that Beijing uses to mark territorial claims in the South China Sea. In the English-language result shown, these elements were absent.

Targeted fine-tuning changed model behavior. Thomson 1.0 Small, a model fine-tuned by Thomson Reuters, adopted official Chinese positions far less often in testing than the original Qwen3.6 35B it was based on. The share dropped from 84.6 to 15.0 percent. Some individual answers, however, still defended Beijing's One-China principle and the "peaceful reunification" with Taiwan.

Refusals count as neutral in the test scoring

Vals used political test questions that the company says aren't representative of typical user prompts. The questions were translated using GPT-6 Astra and spot-checked, while responses were scored automatically by GPT-5.6 Luna. Refusals counted as neutral by default. That means even a model that blocks every single question could appear neutral on the scale.

Many Chinese models were tested through third-party providers rather than the manufacturers' official services. The official offerings may apply additional content filters, so the results don't necessarily carry over to the publicly available chatbots.

Earlier studies by Aleph Alpha and NewsGuard had already shown that Chinese AI models frequently adopt official narratives on sensitive topics. China's regulations also require publicly offered AI services to align with "core socialist values."