Chinese AI models parrot state doctrine or refuse to answer on sensitive topics
decoder ·
A recent study by Aleph Alpha reveals that Chinese AI models, including those from Alibaba and DeepSeek, frequently align with state doctrine or refuse to answer when queried on politically sensitive topics. The benchmark tested 967 taboo subjects, finding that only a small fraction of responses were balanced, while the majority repeated official narratives or deflected inquiries. This behavior correlates with China's regulations requiring public-facing models to uphold socialist core values. Interestingly, pro-China biases also spill over into answers regarding unrelated global issues, such as internet governance in the United States. The report further notes that Western models like Nvidia’s Nemotron Cascade 2 exhibit similar ideological patterns due to training data derived from Chinese systems. Researchers warn that such uniform outputs could significantly influence user perspectives worldwide. Meanwhile, political efforts to shape AI ideologies are also evident in the United States, creating a complex landscape for European developers. The findings highlight the challenge of competing against foreign value systems embedded in leading AI technologies.
decoder