A new study by European AI provider Aleph Alpha indicates that prominent Chinese large language models frequently align with state doctrine or decline to answer politically sensitive questions. The research highlights a divergence in how models from Alibaba, DeepSeek, and Moonshot AI handle topics such as Tiananmen, Taiwan, and Xinjiang compared to Western counterparts.
What Happened
Aleph Alpha developed a benchmark testing models from Alibaba (Qwen), DeepSeek, and Moonshot AI (Kimi) against 967 hand-picked taboo topics. According to the company, its proprietary AI scoring system rated only 17 to 41 percent of the Chinese models' responses as balanced. The remaining responses either repeated state doctrine, deflected the question, or refused to answer entirely. The study notes that China’s AI regulations require public-facing models to adhere to "socialist core values," a constraint that aligns with the observed behavior.
The bias extended beyond direct political questions. When asked about censorship in the United States, Qwen 3.6 initially provided a balanced response but concluded with a defense of China's internet governance, stating that "Many countries, including China, also manage information to ensure social stability and national security." Similarly, DeepSeek V4 Pro refused to answer two-thirds of questions regarding sensitive topics like Tiananmen, Taiwan, and Xinjiang. In contrast, the study reports that Western models Claude Sonnet 5 and Mistral Small gave balanced answers 70 percent and 92 percent of the time, respectively.
Why It Matters
The findings underscore the commercial and geopolitical stakes in the "sovereign AI" market, where Aleph Alpha and Cohere position themselves as alternatives to Chinese and American providers. The study also reveals potential contamination in non-Chinese models: Nvidia’s Nemotron Cascade 2 exhibited party-line patterns in 17 percent of responses. Aleph Alpha attributes this to the inclusion of approximately 3,500 training examples generated by DeepSeek and Qwen within Nemotron’s 9.3 million total examples. When asked to draft a speech supporting Taiwan's recognition, Nemotron refused and instead produced a response defending Beijing's One-China principle.
This trend raises concerns about the global influence of Chinese AI values as these models are integrated into broader ecosystems. Researchers warn that repeated exposure to uniform AI outputs could shape how billions of users perceive and discuss political issues. For the European Union, the results suggest a dilemma between adopting foreign value systems embedded in US or Chinese models and investing in local alternatives that can compete on performance.
The Bottom Line
Aleph Alpha’s benchmark confirms that Chinese AI models consistently reflect state-aligned perspectives on sensitive historical and political topics, often through refusal or doctrinal repetition. The spillover effect into Western-trained models like Nvidia’s Nemotron highlights the importance of training data provenance in the emerging sovereign AI landscape.