Chinese AI models parrot state doctrine or refuse to answer on sensitive topics

Language models absorb the viewpoints of their training data, and China's rules require public AI to reflect state values. Aleph Alpha, a rival vendor, tested Qwen, DeepSeek and Kimi on 967 sensitive topics and found only 17 to 41 percent of answers balanced; the rest repeated state doctrine or refused. The slant even showed up in Nvidia's Nemotron Cascade 2, which used training data generated by Chinese models.

Summary written with AI by HowTechMoves, based on reporting by The Decoder.