Vals study finds Chinese AI models show political complexity, not just government alignment

Vals study tested multiple Chinese and non-Chinese models' alignment with Chinese government positions. The research found that Chinese models overall show lower pro-government lean than expected, with significant variation between individual models. Language has a strong effect on political slant: responses in Simplified Chinese were more pro-government than those in English, Russian, or Spanish. Some models, like GLM-5.3 Flash, explicitly confirmed documented human rights violations, while others, like Kimi K3, redefined Xinjiang detention camps as vocational training centers. The study also found that system prompts indicating mainland China usage increased pro-government responses, and that models sometimes adopted official positions rather than refusing to answer sensitive questions.