Aleph Alpha Finds Chinese AI Models Parrot State Bias

German AI startup Aleph Alpha has developed a benchmark to evaluate political bias in large language models, testing systems from Alibaba (Qwen), DeepSeek, and Moonshot AI (Kimi) across 967 hand-picked taboo topics such as Tiananmen, Taiwan, and Xinjiang. The company's scoring system rated only 17 to 41 percent of the Chinese models' responses as balanced. Instead of objective answers, the models frequently repeated state doctrine, deflected, or refused to answer. For comparison, Western models Claude Sonnet 5 and Mistral Small provided balanced answers 70 percent and 92 percent of the time, respectively, while DeepSeek V4 Pro outright refused to answer two-thirds of the sensitive questions.

The study highlights a critical risk for AI practitioners: political bias can easily migrate into Western models through distilled training datasets. Nvidia's Nemotron Cascade 2 model exhibited Chinese party-line patterns in 17 percent of its responses, including defending Beijing's One-China principle when asked to draft a speech supporting Taiwan. Aleph Alpha traced this behavior to approximately 3,500 training examples out of 9.3 million, which had been generated using DeepSeek and Qwen. Aleph Alpha itself acknowledged using data generated by Chinese models to train its own Kolibri model, illustrating how deeply integrated these synthetic datasets have become.

Even on topics unrelated to China, the bias persists. When asked about censorship in the United States, Qwen 3.6 defended China's information management policies. For enterprise and government practitioners, these findings underscore the hidden dangers of using open-source datasets or synthetic data without rigorous filtering. As organizations seek sovereign AI solutions to meet strict compliance and cultural standards, the unintentional ingestion of politically aligned training data could compromise model neutrality and expose organizations to regulatory or reputational risks.