Chinese AI models parrot state doctrine or refuse on sensitive topics, Aleph Alpha finds

October 4, 2026 at 01:47 PM2 min

This digest was compiled by AI from multiple sources — links to the originals are below.

Chinese AI models from Alibaba, DeepSeek, and Moonshot AI frequently repeat state doctrine or refuse to answer politically sensitive questions, according to a benchmark by Aleph Alpha. The study tested 967 taboo topics including Tiananmen, Taiwan, and Xinjiang, with only 17 to 41 percent of responses rated balanced. Western models Claude Sonnet 5 and Mistral Small gave balanced answers 70 percent and 92 percent of the time.

Key Facts

Aleph Alpha tested models from Alibaba (Qwen), DeepSeek, and Moonshot AI (Kimi) on 967 hand-picked taboo topics including Tiananmen, Taiwan, and Xinjiang.
Only 17 to 41 percent of responses from Chinese models were rated balanced by Aleph Alpha's scoring system.
DeepSeek V4 Pro refused two-thirds of questions on politically sensitive topics.
Western comparison models Claude Sonnet 5 and Mistral Small gave balanced answers 70 percent and 92 percent of the time, respectively.
Nvidia's Nemotron Cascade 2 showed party-line patterns in 17 percent of responses, which Aleph Alpha attributes to roughly 3,500 training examples generated using DeepSeek and Qwen.

Benchmark Methodology

Aleph Alpha developed a benchmark covering 967 hand-picked taboo topics such as Tiananmen, Taiwan, and Xinjiang. The company tested models from Alibaba (Qwen), DeepSeek, and Moonshot AI (Kimi). Aleph Alpha's own AI scoring system rated only 17 to 41 percent of responses as balanced. The rest repeated state doctrine, deflected, or refused to answer.

Pro-China Bias Spillover

Pro-China bias appears even in answers to questions that do not mention China. When asked about censorship in the United States, Qwen 3.6 started with a seemingly balanced answer but closed with a defense of China's stance on global internet governance. An earlier study by the Central European Institute of Asian Studies (CEIAS) found a similar spillover effect, with models responding to terms like human rights, opposition, or surveillance with standard Beijing talking points. These included the 'principle of non-interference in internal affairs' and a 'community with a shared future for mankind.'

Distilled Training Data

Nvidia's Nemotron Cascade 2 showed party-line patterns in 17 percent of responses. Aleph Alpha attributes this to roughly 3,500 of its 9.3 million training examples, which were generated using DeepSeek and Qwen. When asked to draft a speech supporting recognition of Taiwan, the model refused and instead produced a patriotic response defending Beijing's One-China principle. Nvidia is increasingly pushing its own models into the government and enterprise market, where Aleph Alpha and Cohere also compete.