Vals AI tested 25 AI models for their alignment with the positions of the Chinese government. According to The Decoder's description, Chinese models responded to political questions less uniformly than expected.
Researchers also found that Western models sometimes adopted Beijing's official positions. The language of the prompt had the strongest influence on the political tilt of the responses.
This matters for evaluating AI in the international arena: the origin of the model alone does not explain the full nature of its political responses. The available data are presented only as a synopsis of the publication, with no details on methodology and test results.
editorial commentary
Why it matters
A likely practical implication is the need to test models in multiple languages using a single political methodology. The next observable signal will be the detailed sample, questions, and quantitative results published by Vals AI. Significant uncertainty stems from the fact that only The Decoder synopsis is currently available.