A study by Aleph Alpha indicates that Chinese AI models frequently mirror state positions or decline to answer when prompted with politically sensitive topics.
Key facts
- •Aleph Alpha tested 967 taboo topics, finding that Chinese models provided balanced responses only 17% to 41% of the time.
- •DeepSeek V4 Pro refused to answer two-thirds of the sensitive questions included in the study.
- •Nvidia's Nemotron Cascade 2 showed party-line patterns in 17% of responses, linked to training data sourced from Chinese models.
- •Western models Claude Sonnet 5 and Mistral Small achieved balanced response rates of 70% and 92%, respectively.
- •Chinese regulations require public-facing AI models to incorporate socialist core values.
A new study by the AI company Aleph Alpha reports that Chinese large language models frequently echo state doctrine or refuse to answer when presented with politically sensitive topics. The research tested models from Alibaba, DeepSeek, and Moonshot AI across 967 topics, including references to Taiwan, Xinjiang, and the Tiananmen Square protests. These findings align with Chinese regulations that mandate the inclusion of socialist core values in public-facing AI systems.
By the numbers
Performance on Sensitive Topics
Aleph Alpha’s benchmark found that only 17% to 41% of responses from the tested Chinese models were rated as balanced. On specific sensitive subjects, models often defaulted to official Beijing talking points or deflected entirely. For instance, DeepSeek V4 Pro refused to answer two-thirds of the questions posed. In contrast, Western models like Claude Sonnet 5 and Mistral Small provided balanced answers 70% and 92% of the time, respectively.
Spillover and Training Data Influence
The study noted that pro-China biases sometimes appear even in questions not explicitly mentioning China. For example, when asked about U.S. censorship, Alibaba’s Qwen 3.6 pivoted to defend China's information management policies. Aleph Alpha also identified similar patterns in Nvidia’s Nemotron Cascade 2, attributing this to the inclusion of approximately 3,500 training examples generated by DeepSeek and Qwen within its 9.3 million total training samples.
Broader Implications for AI Development
Researchers suggest that language models inherently reflect the cultural and political values present in their training data. While Chinese models are shaped by state requirements, the study notes that ideological shaping also occurs in the United States, citing modifications to Elon Musk's Grok AI. The report concludes that European nations face a choice between foreign value systems unless they can develop competitive, independent AI models.
Advertisement
This article was independently rewritten by ManyPress editorial AI from reporting originally published by The Decoder.
