US-based artificial intelligence research lab Nous Research has retrained Chinese tech giant Alibaba’s popular open-source language model Qwen3.8-27B to remove Beijing’s built-in state censorship. The updated version, dubbed Arete-27B, has been optimized “to reinforce Western values.”
US-based artificial intelligence research lab Nous Research has retrained Chinese tech giant Alibaba’s popular open-source language model Qwen3.8-27B to remove Beijing’s built-in state censorship. The updated version, dubbed Arete-27B, has been optimized “to reinforce Western values.”
Like most Chinese AI models, Qwen3.8-27B comes with built-in restrictions that align with the position of the Chinese Communist Party, Cybernews writes . In the basic version, the model usually refuses to respond to topics sensitive to Beijing or uses euphemisms - for example, regarding the events in Tiananmen Square in 1989, when the military used combat weapons against unarmed students.
Nous Research notes that political bias in language models poses a serious risk today as users increasingly rely on AI agents for fact-finding and decision-making. In addition, ideological biases can have unexpected technical consequences.
“Of course, much of the concern around the models themselves is ‘political’ in nature: the model promotes partisan narratives. But the model’s worldview can have strange, unintended 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 lab explained.
These findings are supported by this year's Booz Allen Hamilton study, which recorded a decrease in the security of code from certain Chinese AIs under certain geopolitical conditions.
After retraining the Arete-27B model, the proportion of responses with elements of state propaganda in tests decreased from 22% to 1%, which is in line with Western counterparts. At the same time, all agent capabilities and basic skills of the model remained unchanged.
The researchers needed about 52.8 million tokens to remove censorship — compared to the scale of the initial model training, the costs for this stage turned out to be negligible. Nous Research noted that the ideological layer in Chinese open models turned out to be surprisingly superficial.
"It's applied late and very thinly, so fifty million tokens of carefully selected supervised learning remove it without affecting anything inside. The values in any open model are always only one retraining away."
