Aleph Alpha, a European AI firm, ran a benchmark that asked Chinese language models 967 questions about politically sensitive subjects. The models tested were Alibaba’s Qwen, DeepSeek and Moonshot AI’s Kimi. Only a small share of their replies were judged balanced; the rest repeated official talking points, deflected, or simply said they could not answer.
How do the Chinese models compare with Western ones?
On the same set of taboo questions, Western models Claude Sonnet 5 and Mistral Small gave balanced answers 70 % and 92 % of the time. DeepSeek’s V4 Pro refused about two‑thirds of the queries, while Qwen and Kimi mostly echoed the party line. Even when the question was not about China, Qwen sometimes slipped a defense of Beijing’s internet policies into its response.
Why the bias appears
China’s AI regulations require public‑facing models to reflect “socialist core values.” That legal backdrop, plus training data that over‑represents state‑approved sources, leads the models to internalise official narratives. Aleph Alpha also noted that Nvidia’s Nemotron Cascade 2 showed party‑line patterns in 17 % of answers, which the company traced to a few thousand training examples that were themselves generated by Chinese models.
Why it matters
For users in China, the bias means chatbots are unlikely to provide neutral information on key political issues, limiting access to diverse viewpoints. Outside China, the tendency can spill over into non‑political answers, subtly shaping perceptions about Chinese policy. The study also shows that regulatory environments can directly steer AI behavior, which matters for anyone choosing a model for public‑facing applications.
