Study Highlights State Doctrine Alignment in Chinese AI Models
A recent study by Aleph Alpha investigates how Chinese language models handle politically sensitive topics, revealing strong adherence to state doctrine.
A recent evaluation conducted by Aleph Alpha has shed light on how major Chinese artificial intelligence models respond to politically sensitive inquiries. The study evaluated models from Alibaba, DeepSeek, and Moonshot AI, focusing on nearly 1,000 hand-picked taboo subjects such as Taiwan, Tiananmen, and Xinjiang.
Aleph Alpha, a company that markets sovereign AI solutions for governments alongside Cohere, utilized its own scoring system to analyze the outputs. According to the findings, only a minority of responses—ranging between 17 and 41 percent—were rated as balanced. The remaining outputs either mirrored official state doctrine, deflected the prompt, or completely refused to answer. These behaviors align with domestic regulations in China that mandate public-facing models to incorporate socialist core values.
For comparison, Western models tested in the benchmark showed higher rates of balanced outputs on the same topics, with Claude Sonnet 5 and Mistral Small providing neutral answers in a clear majority of cases. Meanwhile, specific models like DeepSeek V4 Pro opted to refuse about two-thirds of the sensitive questions.
Spillover Effects in General Queries
The ideological slant is not restricted strictly to direct questions about Chinese politics. Researchers observed that the pro-China bias can occasionally surface in queries that have no initial connection to the country.
For instance, when asked about censorship practices in the United States, Qwen 3.6 initially provided a neutral overview before concluding with a defense of China's approach to global internet governance, arguing that information management is necessary for social stability and national security. On general, non-political topics, however, the tested models largely delivered balanced responses, though faint traces of bias remained visible in systems like Qwen 3.6 and DeepSeek V4 Pro.
Previous research by the Central European Institute of Asian Studies identified similar spillover patterns. When terms related to human rights, opposition, or surveillance appeared in prompts, the systems frequently reverted to standard Beijing talking points, such as the principle of non-interference.
Training Data and Global Implications
The study also examined how these viewpoints can transfer across different systems through training data. Aleph Alpha pointed to Nvidia's Nemotron Cascade 2, which exhibited party-line phrasing in a small percentage of its responses. The researchers attributed this behavior to a tiny fraction of its training dataset that had been generated using DeepSeek and Qwen.
Experts note that language models inherently absorb cultural and political values from their training corpora due to overrepresented viewpoints or intentional curation. This phenomenon raises broader questions about how widespread AI usage might influence public perception globally, especially as political entities in various regions seek to shape model behavior.
What it means for developers
Developers building applications across global markets must account for how different foundational models handle geopolitical nuances, sensitive keywords, and localized compliance mandates. Depending on the target audience, incorporating models with varying underlying training backgrounds can significantly alter output reliability and neutrality. Developers who want to test top AI models cheaply through one API can do so at https://apixoai.online, allowing for straightforward comparisons of response behaviors across multiple frontier systems.
Source: Chinese AI models parrot state doctrine or refuse to answer on sensitive topics — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
