Americans re-educated Alibaba’s AI to say things China usually censors
The project saw a dramatic drop in output that reflects Communist Party of China’s preferred messaging.
AI lab retrains Alibaba's Qwen model. By Cybernews.
Follow us
- Nous Research retrained Alibaba’s Qwen3.8-27B model and renamed it Arete-27B to reduce state-aligned censorship.
- The lab says propaganda-like answers fell from 22% to 1% in its evaluation.
- Researchers say the extra training preserved the model’s task-handling abilities.
- Nous Research argues the model’s China-aligned bias was shallow and could be changed with limited retraining.
Key Takeaways by nexos.ai, reviewed by Cybernews staff.
An American AI lab has subjected Qwen3.8-27B, a popular open-weight LLM from Chinese tech giant Alibaba, to re-education through training in order “to reinforce Western values.”
Like most China-made AI models, Qwen3.8-27B ships with state-aligned behavior baked into the weights. In other words, Beijing makes sure the model’s output is censored.
But Nous Research, a Texas-based AI lab, has decided to re-educate the model so that it tells the user the truth about, say, events during 1989 in Beijing’s Tiananmen Square when soldiers used live ammunition against unarmed students and killed hundreds of them.
According to Nous Research, the censorship embedded in Chinese AI models poses risks to users, as these systems generally decline to address sensitive subjects or rely on euphemistic language.
We post-trained over Alibaba's Qwen3.8-27B to reinforce Western values on a Chinese-trained model. We call this new model Arete-27B,Nous Research said in a post.
All frontier labs have exhibited political bias in the past, but these biases have hard implications now, as users increasingly depend on agents for fact-finding and judgment-heavy work, Nous Research further explained.
“Of course, a lot of the trepidation surrounding the models themselves is ‘political’ in nature: the model pushes party-aligned narratives. But the viewpoint of the model can have strange unintended consequences,” said the lab.
“Several studies found that geopolitical information in the models’ context can affect code quality, with Chinese models producing more vulnerable code when they believe they are acting on behalf of an American entity.”
Bias layer pretty shallow
This year’s study by Booz Allen Hamilton found that some Chinese models produce more vulnerable code under certain conditions.
Stay updated with our latest stories and follow us on social media
Be the first to discover new stories, ideas, and updates from our team.
Nous Research said it observed “firsthand refusals and deflections” and post-trained the Qwen3.8-27B model to alter those biases.
On the lab’s evaluation, Arete-27B cut answers containing propaganda from 22% to 1%. That’s in line with the Western models the lab tested – plus, agentic capabilities were retained.
The fine-tuning required about 52.8 million tokens, the lab added, but explained that against the scale of the model’s pre-training, the cost of the additional post-training was “insignificant.”
Has your password leaked?
Interestingly, Nous Research claims the whole exercise wasn’t difficult. Researchers say they found the bias layer in a Chinese-aligned model to be shallow.
“It is applied late, it is applied thin, and fifty million tokens of carefully chosen supervision removes it without disturbing anything underneath. The values in any open model are only ever one fine-tune away,” said the AI lab.
