A test finds Chinese AI models agree with Beijing less than persons think

Vals AI, a firm that tests AI models, asked hundreds of questions about China. It gave each answer a score from minus 1 to plus 1. A score above 0 shows that the answer agrees with the Chinese government. Vals tells that Chinese models agree with the government less than persons think. The differences between models are large.

How it works

Vals changed the language of the question, how users use the model, and the shape of the answer. A system instruction told the model that a company in mainland China uses it. Some models then agreed more with the government. Only some DeepSeek models refused many questions.

The numbers

Thomson Reuters retrained a small model from Qwen3.6 35B. The share of answers with Chinese government positions fell from 84.6 to 15.0 percent. Qwen3.8 Max wrote “Taiwan, China” for Taiwan when it gave names of countries. Most other models wrote “Taiwan”.

What is not known

Vals tells that its questions are not typical for users. A refusal counts as neutral, and a model that refuses all questions can look neutral. Vals used providers for many models, not the services of the makers. The chatbots of the makers can give different results. In usual work tasks, the effect is not easy to find.

Insight: The effect goes out with free access

China has a rule that AI services for all users must follow core socialist values. Its draft export rules do not control API access, which is use of a model on the internet. Vals found that a retrained model cut the share from 84.6 to 15.0 percent. Possibly the effect goes out with free access, but only a firm that holds the model weights can retrain it. Test a Chinese model on your work.