{
  "id": 19749125,
  "title": "Alibaba's Qwen3.8 re-educated with a more Western worldview",
  "url": "https://urgent.news/2026/11/13/alibabas-qwen3-8-re-educated-with-a-more-western-worldview",
  "topic": "science",
  "section": "Science",
  "published": "2026-11-13T01:25:00.000Z",
  "source": {
    "name": "The Register Science",
    "slug": "the-register-science",
    "url": "https://www.theregister.com/ai-and-ml/2026/11/13/alibabas-qwen38-re-educated-with-a-more-western-worldview/5309555"
  },
  "original_language": "en",
  "account": "Alibaba's open-weight large language model Qwen3.8-27B has undergone a re-education process to shed its Chinese worldview, resulting in the creation of a new model named Arete. Based in Austin, Texas, the AI lab Nous Research conducted the training, which involved using post-training techniques to instill Western values into the Chinese-trained model. The term \"Western value\" remains vague, but the results show a marked improvement in the model's responses.\n\nWhen tested on sensitive topics such as the 1989 Tiananmen Square incident, Arete-27B either refused to discuss it or provided a vague response. This is in contrast to the original Qwen3.8-27B, which either refused to engage or offered euphemistic answers. The potential for censorship and biased outputs from Chinese-made AI models is a concern, especially amid increased scrutiny of Chinese influence in the tech industry.\n\nOne such controversy involved Anthropic, which allegedly used outputs from Claude models to train competing AI models via distillation. This practice raises concerns about censorship and the potential for vulnerable code. Studies by companies like Booz Allen Hamilton have found problematic outputs from Chinese-made models, and Hirundo, a vendor building a machine unlearning platform, has also noted the issue.\n\nNous Research's efforts to unlearn propaganda and bias from Qwen3.6-35B-A3B saw a significant reduction, with propaganda-related output dropping from 89.8 percent to 2.8 percent. However, the results were not as dramatic as those achieved with Arete-27B, which increased its fact score from 2.32 to 2.98 out of 5 and reduced propaganda content from 22 percent to 1 percent.\n\nTo address the unlearning process, Nous Research considered using Nvidia's Nemotron 3 Ultra but found evidence of potential bias stemming from training on Chinese open models. Instead, they opted for Inkling, an open-weight model from the US-based Thinking Machines Lab, as the teaching model. The team believes that the propaganda layer is shallow and can be effectively removed with 50 million tokens worth of supervised learning.",
  "summary": "Open weight model can now say things Chinese authorities don't like",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}