A study by Vals AI examined the political alignment of Chinese and Western language models. Researchers used a scale from -1 to +1, where positive values indicate alignment with Chinese government positions and negative values reflect critical responses.
The findings show that Chinese models are less uniform than previously assumed. While some models adopted official narratives, others provided critical answers. For instance, GLM-5.3 Flash acknowledged documented human rights violations in China.
Prompt language significantly influences political bias. Responses in Simplified Chinese were generally more pro-government than those in English, Russian, or Spanish. This effect was observed in both Chinese and American models.
System instructions and task formats also impact results. Models often shifted toward official positions when prompted to answer with simple yes or no responses, or when the system prompt suggested the model was being used within mainland China.
Fine-tuning can reduce pro-government bias. The Thomson 1.0 Small model, based on Qwen3.6 35B, showed a decrease in alignment with official Chinese positions from 84.6 percent to 15.0 percent after targeted fine-tuning.
