Nous Research has made significant strides in addressing political bias in artificial intelligence by retraining Alibaba’s Qwen3.8-27B model. The result of this effort is Arete-27B, which has successfully reduced the model’s pro-Chinese propaganda from 22 percent to less than one percent.
This transformation was achieved through a method involving 53 million tokens of supervised fine-tuning, demonstrating that the political bias present in the original model was merely a ‘thin layer’ that could be stripped away. Testing has shown that the general capabilities of the AI remain intact, highlighting the potential for enhancing AI models without compromising their functionality.
The findings underscore the importance of transparency and ethical considerations in AI development, particularly in models that may be influenced by state propaganda. As AI continues to evolve, the ability to refine and improve existing models while mitigating biases will be crucial for their responsible deployment in various applications.
Compiled automatically by the Tech AI Newsdesk from public AI-news sources and summarised in our own words.
