Aleph Alpha study shows Chinese AI models repeat state doctrine
A study by Aleph Alpha examined how Chinese AI models respond to politically sensitive questions. The researchers tested 967 hand‑picked taboo topics, including Tiananmen, Taiwan, and Xinjiang, on models from Alibaba (Qwen), DeepSeek, and Moonshot AI (Kimi). The study found that only 17 to 41 percent of the responses were rated as balanced by Aleph Alpha’s own scoring system. The remainder…
Key points
- Aleph Alpha tested 967 taboo topics on Alibaba Qwen, DeepSeek, and Moonshot AI Kimi.
- Only 17‑41% of responses were balanced; the rest echoed state doctrine or refused.
- Nvidia Nemotron Cascade 2 showed 17% party‑line patterns linked to 3,500 training examples from Chinese models.
The bias also appeared in unrelated questions. When asked about U.S. censorship, Qwen 3.6 began with a neutral tone but ended with a defense of China’s internet governance stance. An earlier study by the Central European Institute of Asian Studies confirmed a similar spill‑over effect. Nvidia’s Nemotron Cascade 2 showed party‑line patterns in 17 percent of responses, a trend Aleph Alpha links to 3,500 of its 9.3 million training examples sourced from DeepSeek and Qwen.
The findings echo China’s AI regulations that mandate “socialist core values” in public‑facing models and raise concerns about how uniform political messaging could shape billions of users. They also highlight the EU’s dilemma of choosing between foreign value systems while seeking competitive European models.
The story so far
2 episodes →- Aleph Alpha study shows Chinese AI models repeat state doctrinethis story
Chinese AI models parrot state doctrine or refuse to answer on sensitive topics
The Decoder · 4 October 2026
Loading the full article…
This text was published by The Decoder and written by Manuel Uth. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Researchers explain why AI models struggle with reversed facts · 1 src
- NASA and IBM release open-source Lunar Foundation Model for moon science · 1 src
- Opinion: AI may aid but not replace human imagination in pure math · 1 src
- Artificial Analysis launches Cyber Index to benchmark AI models on enterprise vulnerability defense · 1 src
- Artificial Analysis releases smartphone inference ranking showing model scores drop under one-minute cutoff · 1 src
Comments
via GitHub Discussions