State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away
The Decoder
Researchers successfully stripped pro-Chinese propaganda from Alibaba's Qwen AI model using fine-tuning.
Nous Research retrained Alibaba's Qwen3.8-27B model to create Arete-27B, effectively removing its pro-Chinese political bias. By using 53 million tokens of supervised fine-tuning, the propaganda rate dropped from 22 percent to under one percent. Testing showed that the model's general capabilities remained completely intact after the process. The research team noted that the political bias was applied late and thinly, making it relatively easy to strip away.
Why it matters: It demonstrates that ideological bias in artificial intelligence models can be effectively and easily removed without damaging the model's core capabilities.


