Making AI Headlines · Page 7
Researchers Strip Political Bias From a Model and Reveal a Very Thin Coat
· New York time
Nous Research says it retrained Alibaba's Qwen3.8-27B model to create Arete-27B, largely removing its pro-Chinese bias. In the tests described, the propaganda rate fell from 22 percent to less than one percent, while general capabilities remained intact. The work used 53 million tokens of supervised fine-tuning. The researchers describe the political bias as a thin layer applied late in the process. Their result concerns the model and tests they examined, rather than a guarantee that every political bias can be removed as easily. It does show how much a model's responses can change through additional training without necessarily losing the capabilities measured in those tests. For developers building on open models, that makes the choice of training data and evaluation an important part of the product rather than a cosmetic afterthought. A model can arrive sounding authoritative, then acquire a notably different voice after someone changes the training. The political message may look like a permanent feature of the building, but this experiment suggests that, in this case, it was closer to the paint. That still leaves a consequential question: who gets to choose the next color?
Read this story in the full newspaper →Satire. Not factual reporting. Read the original reporting from the linked publishers.