Researchers evaluated the political bias of models from China

The company Vals AI has published the results of testing 25 language models for their alignment with the political line of the Chinese authorities. The authors of the study checked how uniformly the systems respond to sensitive questions.

The research revealed that Chinese models demonstrate much greater variability in their responses than initially assumed. At the same time, Western neural networks also reproduced Beijing's official wording in a number of cases.

A key factor influencing the ideological bias of the responses turned out to be the language of the query. Changing the prompt language significantly altered the tone and political coloring of the generated content.

The findings help developers and researchers better understand the mechanisms of censorship and the influence of language barriers on the safety of AI systems in various geopolitical contexts.