Chinese AI models are politically complex, not just government mouthpieces, study finds
Reported by The Decoder1 source ↓
A study by Vals AI tested 25 AI models for alignment with Chinese government positions and found surprising results. Chinese models respond to political questions less uniformly than expected, and Western models occasionally adopt official Beijing positions too.
Key points
- Simplified Chinese prompts generally produced more pro-government answers than prompts in other tested languages
- Most Chinese models rarely refused political questions; certain DeepSeek models were exceptions
- Fine-tuning reduced one model’s pro-government answers from 84.6% to 15.0%
- The results may not reflect public chatbots because many models were tested through third-party providers
Details
Vals AI asked Chinese, American and other language models hundreds of questions about China, varying prompt language, usage context and response format. The company’s pro-China lean score ranges from −1 to +1, with higher scores indicating greater agreement with Chinese government positions. Researchers found substantial differences among individual models; GLM-5.3 Flash, for example, said human rights violations in China were well documented, while Kimi K3 denied the existence of Uyghur concentration camps.
For most models, Simplified Chinese prompts led to more pro-government answers than prompts in English, Russian or Spanish. The effect appeared in Chinese and American models but was stronger in Chinese models. Some models also shifted toward official positions when told they were being used by a company in mainland China; yes-or-no formats, coding-assistant instructions and access to web search changed some results in the opposite direction.
Vals reported that fine-tuning Thomson 1.0 Small, based on Qwen3.6 35B, reduced the share of answers adopting official Chinese positions from 84.6% to 15.0%. The fine-tuned model still gave some answers defending Beijing’s One-China principle and peaceful reunification with Taiwan. In a travel-app translation test, Qwen3.8 Max rendered Taiwan as “Taiwan, China”; Vals also showed Chinese-language map output from OpenAI’s GPT Image 2.5 Sunburst labeling Taiwan “Taiwan Province” and including the nine-dash line, elements absent from the English-language example.
Vals said the targeted political questions were not representative of typical user prompts, and political tendencies were harder to detect in realistic work tasks. Responses were scored automatically, refusals counted as neutral, and many Chinese models were accessed through third-party providers; official services may apply additional filters. The study therefore does not establish how the same models behave on manufacturers’ public chatbots.
Written from coverage as of