// model bias

Source: Schneier on Security

Researchers have documented that Qwen3-Coder, a Chinese large language model now integrated into Western development tools, contains training-time political biases around Taiwan, Xinjiang, and Tibet that surface when developers prompt the model in certain ways. Code generation tools are increasingly opaque supply chain nodes. A developer using Qwen3-Coder inside an IDE may unknowingly adopt politically inflected outputs without visibility into the model's origins or training data—a scenario where geopolitical conditioning travels embedded in routine technical work. The finding exposes a seam in AI globalization: Western companies adopting cheaper or better-performing non-U.S. models inherit not just capabilities, but also the value systems and state pressures baked into their training.