Chinese AI models echo state doctrine on sensitive topics
aleph alpha study finds chinese ai models like alibaba's qwen and deepseek toe the party line on 967 taboo topics, repeating state doctrine or refusing to answer, affecting users and raising concerns about biased information.
aleph alpha, a provider of 'sovereign ai' for governments, conducted a study on chinese ai models from alibaba (qwen), deepseek, and moonshot ai (kimi) using a benchmark of 967 hand-picked taboo topics. the models frequently parroted state doctrine or refused to answer, with only 17-41% of responses rated as balanced. deepseek v4 pro refused to answer two-thirds of questions on sensitive topics. western models like claude sonnet 5 and mistral small gave balanced answers 70% and 92% of the time, respectively. the study also found that chinese models can inject pro-china bias into unrelated answers. aleph alpha attributes this bias to china's ai regulations and training data.