Study: Chinese AI models less aligned with Beijing than expected

A Vals AI study evaluates how 25 AI models align with Chinese government political positions. Findings reveal unexpected diversity: Chinese models respond less uniformly to sensitive questions than anticipated, and some Western models occasionally adopt Beijing's official stances.

Researchers tested Chinese, American, and other models on hundreds of political questions, adjusting language, context, and response formats. Alignment with Chinese positions, measured on a scale from −1 to +1, shows Chinese models have a weaker overall pro-government bias than expected. Some, like GLM-5.3 Flash, explicitly acknowledge human rights violations in China. Conversely, non-Chinese models sometimes adopt official Chinese positions. The prompt's language strongly influences bias: responses in Simplified Chinese are more pro-government than in English, Russian, or Spanish, with a stronger effect for Chinese models.

These differences persist across contexts such as agentic frameworks or external tool access, where responses become less pro-government. However, political tendencies are harder to detect in realistic tasks than in targeted tests. The study also highlights the impact of system prompts and response formats, as well as variability between models. It notes that content filters in official releases could alter these behaviors.