A study by Aleph Alpha analyzed how Chinese AI models handle politically sensitive topics. The research tested models from Alibaba, DeepSeek, and Moonshot AI using a benchmark of 967 taboo subjects, including Taiwan and Xinjiang.
The analysis found that only 17 to 41 percent of responses were balanced. The remaining outputs either mirrored state doctrine, deflected the question, or refused to provide an answer, aligning with Chinese regulations on socialist core values.
The study also identified a spillover effect where models introduced pro-China narratives even when asked about unrelated topics, such as censorship in the United States. Similar findings were previously reported by the Central European Institute of Asian Studies.
Aleph Alpha also examined Nvidia's Nemotron Cascade 2, noting that 17 percent of its responses followed party-line patterns. The company attributed this to the inclusion of roughly 3,500 training examples generated by DeepSeek and Qwen within a dataset of 9.3 million entries.
