The American artificial intelligence research laboratory Nous Research conducted additional training of the popular open language model Qwen3.8-27B from the Chinese tech giant Alibaba in order to remove Beijing's built-in state censorship. The updated version was named Arete-27B and was optimized "to reinforce Western values."
Why developers are removing political filters
At Nous Research, they emphasize that political bias in language models today poses a serious risk, as users increasingly rely on AI agents for fact-finding and decision-making. In addition, ideological adjustments can have unexpected technical consequences.
"Of course, a significant portion of the concern around the models themselves is 'political' in nature: the model promotes party narratives. But a model's worldview can have strange, unpredictable consequences. Several studies have found that geopolitical information in the context of models can affect code quality — Chinese models generate more vulnerable code when they believe they are acting on behalf of an American entity," the laboratory explained.
These conclusions are also confirmed by this year's study by Booz Allen Hamilton, which recorded a decline in code security from certain Chinese AIs under specific geopolitical conditions.
Results of "retraining" and the ease of hacking censorship
After additional training of the Arete-27B model, the share of responses containing elements of state propaganda in tests decreased from 22% to 1%, which corresponds to the indicators of Western analogues. At the same time, all agentic capabilities and basic skills of the model remained unchanged.
It took researchers about 52.8 million tokens to eliminate the censorship — compared with the scale of the model's original training, the costs of such a stage turned out to be negligible. Nous Research noted that the ideological layer in Chinese open models turned out to be surprisingly superficial.
"It is applied late and in a very thin layer, so fifty million tokens of carefully curated supervised learning remove it without affecting anything inside. The values in any open model are always only one fine-tuning away."