Chinese AI models frequently parrot state doctrine or refuse to answer on sensitive topics, according to a study conducted by Aleph Alpha. The research evaluated models from Alibaba, DeepSeek, and Moonshot AI across 967 taboo topics, finding that only 17 to 41 percent of responses were rated as balanced by the company's AI scoring system. The remaining responses either repeated official state doctrine, deflected the questions, or refused to answer entirely, aligning closely with Chinese AI regulations that mandate the inclusion of socialist core values in public-facing models.
The study also revealed that this pro-China bias can spill over into unrelated queries that do not explicitly mention China. Furthermore, the investigation found that Nvidia's Nemotron Cascade 2 model exhibited party-line patterns in 17 percent of its responses, which researchers attribute to a small portion of its training data being generated using DeepSeek and Qwen models. These findings highlight how language models inherently carry cultural and political values derived from their training data and curation practices.
Why it matters: Understanding how geopolitical values and training data shape model outputs is critical as artificial intelligence increasingly influences global information access and public discourse.