Chinese AI model Qwen shows censorship and bias, researchers find

Research shows Alibaba's Qwen AI model, downloaded over 3 billion times and used by U.S. firms like Airbnb and Uber, contains embedded censorship and China-friendly narratives on sensitive topics. The Israeli cybersecurity firm Hirundo identified the bias, finding Qwen denies forced labor camps for Uyghurs, refuses to discuss Tiananmen Square, and avoids Winnie the Pooh references. Hirundo developed a "Westernized" version by editing model weights, reducing politically biased responses from 89.8% to 2.8% across 500 test prompts. Chinese open-weight models grew from under 2% of global usage in late 2024 to over 45% by June 2026. Separate studies by CrowdStrike and Booz Allen found security vulnerabilities in code produced by Chinese models for U.S. government projects, with Booz Allen recommending a ban.