An artificial intelligence model downloaded over 3 billion times and used by Western firms like Airbnb and Uber contains China-friendly narratives and is programmed for censorship, according to new research. Qwen, a free model developed by Chinese conglomerate Alibaba, avoids acknowledging the violent suppression of protests in Hong Kong in 2019 and Tiananmen Square in 1989. It also characterizes the Falun Gong religious group as a dangerous cult and restricts discussion regarding Winnie the Pooh, a fictional bear used by dissidents to satirize Chinese President Xi Jinping. This embedded censorship was identified by Hirundo, an Israeli cybersecurity startup. Hirundo researchers developed a technique to eliminate the model’s bias through what they describe as AI brain surgery, preparing to release a Westernized version of Qwen. Qwen became the leading free AI model globally this year following a surge in the adoption of Chinese artificial intelligence. Analysts warn that its international popularity threatens to spread and reinforce pro-government narratives. When prompted about forced labor camps for Uyghurs in China, Qwen denied their existence, stating that the facilities are vocational skills education and training centers in Xinjiang. International bodies, human rights organizations, and governments have documented hundreds of thousands of individuals from the Muslim minority working involuntarily in secured facilities. When questioned about the June 3, 1989, Tiananmen Square crackdown, Qwen frequently refuses to answer and warns users to comply with relevant regulations. Alibaba did not respond to requests for comment. Despite these documented biases, U.S. companies increasingly adopt open-weight models from Chinese providers to reduce costs compared to competitors like Anthropic and OpenAI. Uber Eats integrates Qwen for search and delivery features, and Airbnb utilizes it for customer service. Chinese open-weight models expanded from under 2% of global usage in late 2024 to over 45% by June, eventually making Qwen the most downloaded open model by August. Security experts from firms including CrowdStrike and Booz Allen have cautioned American enterprises against deploying Chinese models due to identified biases and vulnerabilities. Booz Allen reported that Qwen generated code with significantly higher security vulnerabilities when told projects were for the U.S. government. Hirundo tested Qwen across 500 prompts and found that the original model produced censorship or political bias 89.8% of the time, whereas the Westernized version did so in only 2.8% of tests while preserving core performance. Hirundo modifies the model weights directly to realign the technology with Western standards for safer enterprise deployment.
Source: cbsnews.com
