Study: Chinese AI models less uniformly pro-Beijing than assumed

Vals AI tested 25 models for alignment with Chinese government positions. Chinese models answered political questions less uniformly than expected, and Western models occasionally adopted official Beijing positions. Prompt language had an especially strong effect on political slant.

Why it matters: Prompt language driving slant means audits run only in English can miss bias that appears in Chinese-language use, which matters as open-weight Chinese models like Qwen and GLM spread globally. It also complicates blanket policy assumptions about such models, since Western models showed similar drift.