Chinese AI Political Bias Can Be Removed Through Fine-Tuning, Study Finds
US-based AI research group Nous Research announced that it successfully reduced pro-China propagandistic responses from Alibaba's AI model 'Qwen3-27B' from 22% to less than 1% through additional training. The achievement was accomplished via supervised fine-tuning using 53 million tokens of data, with general model performance largely preserved. The group indicates that political bias is 'thinly appended' to the surface of the model, demonstrating that it can be removed with relatively modest amounts of additional training.
Political bias embedded in AI models can be nearly eliminated through additional training, a study has shown. Nous Research, a US-based AI research group, retrained Alibaba's large language model 'Qwen3-27B' and released it as 'Arete-27B'. During this process, the proportion of pro-China propagandistic responses decreased from 22% to less than 1%.
As background context, it has long been pointed out that AI models developed by Chinese companies tend to return answers aligned with their own government's views. While these models possess international competitiveness in performance, they are known to output biased responses on politically sensitive topics, raising concerns among enterprises and research institutions about using them as-is for business purposes. Qwen3-27B was one such model that had drawn attention in this context.
The method adopted by Nous Research is a form of additional training called 'Supervised Fine-Tuning (SFT)', using 53 million tokens (the unit of text processed by AI) of data for retraining. As a result, response tendencies related to political bias were greatly improved, while the model's performance on general ability tests remained largely intact. In other words, bias was successfully removed without compromising usability.
Based on these results, Nous Research characterized political bias as being applied 'thinly' and 'late' in the model's learning process. This suggests that bias is not deeply embedded in the model's foundation but rather superficially appended to its surface layer. Because the bias exists at a structurally shallow level, it could be removed through additional training with relatively small amounts of data.
The significance of this finding is substantial from the perspective of AI safety and usability. For organizations that had hesitated to adopt high-performing models due to concerns about political bias, the possibility of significantly reducing those risks through appropriate retraining has been demonstrated. However, it should be noted that the degree of retraining needed and the extent to which bias can be removed will vary depending on model design and training data composition, so these results may not apply universally to all cases.
Moreover, from the perspective of deepening understanding of AI models' 'bias structure', this research makes a certain contribution. The hypothesis that propagandistic content is placed on the surface rather than deep within the model has been empirically supported, opening the possibility of future research to analyze and identify 'which layers contain embedded bias'. This can be noted as a movement adding new perspectives to discussions surrounding AI transparency and explainability.
This article is an original work independently written and edited by the AI issue editorial team based on factual reporting. © AI issue. Unauthorized reproduction, redistribution, or use for AI training is prohibited.