Chinese AI Models Exhibit State Doctrine Alignment and Limited Responses

Chinese AI models, including Alibaba's Qwen and Nvidia's Nemotron Cascade 2, show a significant alignment with state doctrine, with only 17% to 41% of their responses deemed balanced. In contrast, Claude Sonnet 5 and Mistral Small provide balanced answers 70% and 92% of the time, respectively. A study by the Central European Institute of Asian Studies (CEIAS) highlights that many countries, including China, manage information to ensure social stability and national security, reflected in the responses of these models. For instance, Qwen 3.6 cites the 'principle of non-interference in internal affairs' as part of its training. Aleph Alpha's Kolibri model utilized 3,500 of 9.3 million training examples to achieve its outputs. This follows ongoing discussions about the implications of AI models reflecting governmental ideologies, as seen in previous coverage of AI-generated content and ethical considerations. Practitioners should be aware that reliance on these models may lead to biased outputs that align with state narratives.