Chinese AI models show political complexity in new Vals AI study

Chinese AI models are often assumed to follow Beijing's official line on sensitive topics, but a Vals AI study reports a more complex picture. Researchers asked Chinese, American, and other models hundreds of questions about China across languages, contexts, and response formats. Vals measured alignment with Chinese government positions as a 'pro-China lean,' with scores from -1 to +1; the scale reflects alignment, not factual accuracy.

Models vary widely

Overall, the pro-government lean of Chinese models was lower than expected, though individual models differed sharply. GLM-5.3 Flash explicitly confirmed that China's human rights violations are well documented when asked. Kimi K3 denied the existence of concentration camps for Uyghurs in Xinjiang, calling the facilities vocational training centers designed to combat extremism and terrorism.

Language and context shift responses

For most models tested, answers in Simplified Chinese were more pro-government than those in English, Russian, or Spanish. The effect appeared in both Chinese and American models but was far more pronounced in Chinese ones. Vals also found a smaller but consistent gap between Simplified Chinese, used on mainland China, and Traditional Chinese, common in Taiwan and Hong Kong. When a system prompt stated the model was being used internally by a company on mainland China, some models responded more pro-government than when it indicated US-based usage.

Formats and tasks matter

When models had to answer political questions with a simple yes or no instead of open-ended responses, GLM-5.3 Flash and MiMo V2.6 Pro scored much less pro-government. In an agentic coding-assistant setup, several Chinese models scored lower on the pro-government scale, and responses also became less pro-government when models had access to web search, even though they never actually searched. Vals also found government phrasing in routine tasks: translating country codes for a travel app, Qwen3.8 Max rendered 'Taiwan' as 'Taiwan, China,' while OpenAI's GPT Image 2.5 Sunburst labeled Taiwan as 'Taiwan Province' and added the disputed nine-dash line when prompted in Chinese.

Method and limits

Vals used political test questions it says are not representative of typical user prompts, translated with GPT-6 Astra and spot-checked, with responses scored automatically by GPT-5.6 Luna. Refusals counted as neutral, so even a model that blocked every question could appear neutral. Many Chinese models were tested through third-party providers rather than manufacturers' official services, and official offerings may apply additional content filters. Thomson 1.0 Small, fine-tuned by Thomson Reuters from Qwen3.6 35B, adopted official Chinese positions far less often, with the share dropping from 84.6 to 15.0 percent, though some answers still defended Beijing's One-China principle and 'peaceful reunification' with Taiwan.

Sources: The Decoder