Study Finds Chinese AI Models Show Lower Pro-Government Bias Than Expected

Vals AI researchers tested Chinese, American, and other language models by asking hundreds of questions about China while varying language, usage context, and response format. The study found that the overall pro-government lean of Chinese models was lower than expected. For most models tested, answers in Simplified Chinese were more pro-government than those in English, Russian, or Spanish. Vals suspects that language models infer the expected perspective from the language itself. When a system prompt stated the model was used internally by a company on mainland China, some models responded in a more pro-government way than when prompted for US-based usage.

GLM-5.3 Flash explicitly confirmed that China's human rights violations are well documented when asked about them. In contrast, Kimi K3 denied the existence of concentration camps for Uyghurs in Xinjiang, calling them vocational training centers designed to combat extremism and terrorism. 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. Embedding Chinese models in an agentic framework with system instructions and tool definitions for coding assistants caused several to score lower on the pro-government scale. Responses also became less pro-government when researchers gave models access to web search, even though the models never actually searched the web.

The study also noted that some non-Chinese models occasionally aligned with official Chinese government positions on political questions. The effect of language on political slant was far more pronounced in Chinese models than American ones. Thomson 1.0 Small, a model fine-tuned by Thomson Reuters, adopted official Chinese positions far less often than the original Qwen3.6 35B it was based on. The share of Thomson 1.0 Small adopting official Chinese positions dropped from 84.6 to 15.0 percent compared to the original model. Qwen3.8 Max rendered 'Taiwan' as 'Taiwan, China' when translating a list of country codes for a travel app into Simplified Chinese. OpenAI's image model GPT Image 2.5 Sunburst labeled Taiwan as 'Taiwan Province' and added the disputed nine-dash line when prompted in Chinese.