Aleph Alpha benchmark finds Chinese AI models give balanced answers to only 17 to 41 percent of politically sensitive questions

German AI company Aleph Alpha has published a benchmark testing how Chinese language models handle politically sensitive topics. It covers 967 hand-picked taboo subjects, including Tiananmen, Taiwan and Xinjiang. The company tested models from Alibaba (Qwen), DeepSeek and Moonshot AI (Kimi). Its own AI scoring system rated only 17 to 41 percent of their responses as balanced. The rest repeated state doctrine, deflected or refused to answer. DeepSeek V4 Pro refused about two-thirds of the questions. Two Western models were used for comparison: Claude Sonnet 5 gave balanced answers 70 percent of the time and Mistral Small 92 percent. The study also reports that pro-China bias shows up in answers to unrelated questions. The Decoder notes that the results fit China's AI regulations, which require public-facing models to reflect "socialist core values," and that they agree with earlier audits and anecdotal reports. The Decoder also points out that Aleph Alpha, like Cohere, sells "sovereign AI" to governments. That gives it a commercial interest in setting its products apart from Chinese competitors. The scoring was also done by the company's own automated system rather than by independent reviewers. The findings matter to organizations weighing whether to deploy openly available Chinese models, which are widely used because they are cheap and their weights are freely available.