Research reveals embedded censorship in Alibaba's Qwen AI, prompting calls for 'Westernized' model edits

Key points

• Alibaba's Qwen model, downloaded over 3 billion times, is programmed to deny facts regarding Uyghur forced labor, Tiananmen Square, and Hong Kong protests, aligning with Chinese government censorship.
• U.S. companies, including Uber and Airbnb, have adopted Qwen for its low cost and open-weight availability, despite warnings from security firms like CrowdStrike and Booz Allen about security vulnerabilities and bias.
• Israeli startup Hirundo claims to have created a 'Westernized' version of Qwen by editing model weights to remove bias, reducing censorship-aligned responses from 89.8% to 2.8% in tests.
• Chinese open-weight models grew from under 2% to over 45% of global AI usage by June 2026, raising concerns about the spread of state-aligned narratives in international tech infrastructure.

Background

This development follows earlier U.S. federal accusations that Chinese AI firms are distilling American models to accelerate their own development, as reported in September 2026. It also aligns with broader global anxieties about AI's impact on jobs and the U.S. government's recent push to accelerate domestic AI growth through new executive initiatives.

How outlets are covering it

CBS News and Hirundo emphasize the direct evidence of censorship in Qwen and the potential of 'AI brain surgery' to mitigate it. Vision Times highlights that censorship is not limited to Chinese models but affects global models like ChatGPT and Gemini when queried in Chinese, suggesting that training data from state-controlled sources permeates all systems. While CBS focuses on the specific risks to U.S. firms using Qwen, Vision Times argues that the issue is systemic, rooted in the dominance of state narratives in Chinese-language data, making language a critical factor in AI bias.

Why it matters

The widespread adoption of Chinese AI models in Western enterprises raises significant security and ethical concerns. If these models embed state-aligned narratives, they could influence public opinion and decision-making in ways that conflict with Western values. The debate over whether to ban, regulate, or 'de-bias' these models will likely shape future AI policy and international tech competition.

What to watch

Hirundo plans to publish its 'Westernized' version of Qwen, while U.S. firms may face increased pressure to audit their AI systems for bias and security flaws. Regulatory bodies may consider restrictions on the use of Chinese open-weight models, and further research will be needed to determine the long-term effects of 'AI brain surgery' on model performance.