Alibaba's Qwen AI models found to have embedded censorship

Research indicates that Alibaba Cloud’s Qwen AI model family has censorship embedded directly into its core training rather than through an external filtering layer. Studies show the models appear to understand sensitive topics—such as Tiananmen Square or Winnie the Pooh—but are trained to deflect or evade these subjects by citing “illegal information.”

This behavior is attributed to the model's training protocols, specifically supervised fine-tuning and reinforcement learning from human feedback (RLHF), which align the AI with Chinese regulations requiring the upholding of core socialist values. Because Qwen is an open-weight model, its embedded biases have been integrated into various use cases at companies globally, including in the United States.

In response, developers at Lazarus AI have introduced the “ReAligned” series, which aims to fine-tune Qwen models to produce more neutral outputs while maintaining their original functional capabilities.