Aleph Alpha evaluated 12 Chinese open‑weight LLMs and found they answered only 17‑41 % of politically sensitive prompts in a balanced way. The test also flagged NVIDIA’s Nemotron Cascade 2 on 17 % of prompts because about 3.5 k of its 9.3 M training rows contain Chinese Communist Party talking points. The company says future models will screen for such content, add targeted alignment data, and use benchmarks like this one.
Why it matters
Understanding political bias in AI helps users and regulators decide how safe and neutral these models are for everyday use.
