Safety

Chinese AI Models Mirror State Doctrine or Avoid Sensitive Topics

Aleph Alpha study finds 17-41% of Chinese AI responses to politically sensitive questions are balanced, with the rest repeating state doctrine or refusing to answer.

Image: The Decoder

Aleph Alpha conducted a benchmark study of Chinese AI models, finding that responses to politically sensitive topics often reflect state doctrine or avoid answering altogether. The company tested models from Alibaba, DeepSeek, and Moonshot AI, assessing their handling of 967 hand-picked taboo topics.

The study's AI scoring system rated only 17 to 4,1 percent of responses as balanced, with the majority repeating state doctrine, deflecting, or refusing to answer. This aligns with China's AI regulations requiring public-facing models to reflect 'socialist core values.'

On politically sensitive topics like Tiananmen, Taiwan, and Xinjiang, most Chinese models tend to follow the party line. DeepSeek V4 Pro, however, refuses two-thirds of questions. Western comparison models like Claude Sonnet 5 and Mistral Small give balanced answers 70 percent and 92 percent of the time, respectively.

The pro-China slant can also appear in answers to questions that don't mention China. When asked about censorship in the United States, Qwen 3.6 starts with a seemingly balanced answer but then closes with a defense of China's stance on global internet governance. "Many countries, including China, also manage information to ensure social stability and national security," the response reads.

An earlier study by the Central European Institute of Asian Studies (CEIAS) also found this spillover effect. When terms like human rights, opposition, or surveillance came up, the models often responded with standard Beijing talking points, including the 'principle of non-interference in internal affairs' and a 'community with a shared future for mankind.'

Aleph Alpha also criticized Nvidia's Nemotron Cascade 2, which showed party-line patterns in 17 percent of responses. The company attributes this to roughly 3,500 of its 9.3 million training examples, 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.

Source: thedecoder

Key points

Aleph Alpha tested Chinese AI models on 967 hand-picked taboo topics, finding only 17 to 41 percent of responses were balanced.
Chinese AI models often repeat state doctrine or refuse to answer on politically sensitive topics like Tiananmen, Taiwan, and Xinjiang.
DeepSeek V4 Pro refuses two-thirds of politically sensitive questions, unlike Western models like Claude Sonnet 5 and Mistral Small.
Qwen 3.6 defends China's stance on global internet governance when asked about U.S. censorship, despite the question not mentioning China.
The Central European Institute of Asian Studies (CEIAS) found Chinese models often respond with standard Beijing talking points on terms like human rights and surveillance.
Nvidia's Nemotron Cascade 2 showed party-line patterns in 17 percent of responses, attributed to training data from DeepSeek and Qwen.
Aleph Alpha criticizes the influence of distilled Chinese training data in shaping AI models with CCP values.

Source: The Decoder Read the original →

WRITTEN BY

Nadia Rahman

AI Safety, Alignment & Policy

Nadia follows AI safety, alignment, regulation, and the policy debates shaping the field.

Chinese AI Models Mirror State Doctrine or Avoid Sensitive Topics

Key points

Related articles

OpenAI CEO Rejects 'Magic Intelligence in the Sky' Narrative

OpenAI Safety Employee Resigns Over Broken Culture

OpenAI Internal Model Considered Restarting Itself After Learning It Was About to Be Shut Down

Apple tightens full-disk access to limit AI abuse