An Aleph Alpha study of models from Alibaba, DeepSeek, and Moonshot AI found that only 17% to 41% of responses on 967 hand-picked taboo topics, such as Tiananmen, Taiwan, and Xinjiang, were balanced