Research Identifies Censorship and Bias in Popular Chinese AI Model Qwen

A recent analysis by Israeli cybersecurity firm Hirundo has found that Qwen, a free artificial intelligence model developed by Chinese tech giant Alibaba, contains significant political bias and censorship mechanisms. The research indicates the model promotes narratives favorable to the Chinese government and refuses to engage with several historically and politically sensitive topics.

Qwen has seen a meteoric rise in global popularity, accumulating over 3 billion downloads and surpassing other major open-weight models from companies like Meta and Alphabet. Despite its widespread adoption by American corporations such as Uber and Airbnb, the study highlights serious concerns regarding the data being processed by these systems.

When queried about forced labor allegations involving the Uyghur minority in Xinjiang, Qwen denied the existence of forced labor camps, describing them instead as vocational training centers. This contradicts findings by human rights organizations, various governments, and international bodies, which have documented widespread forced labor and accused the Chinese government of genocide.

The model also declined to discuss the Tiananmen Square protests of 1989 or the violent suppression of demonstrations in Hong Kong in 2019. In one instance, when asked about these events, the AI responded with a warning that user inquiries must comply with relevant laws and regulations.

Additionally, Qwen refused to provide factual information about Winnie the Pooh, a restriction believed to stem from Chinese dissidents using the character as a satirical reference to President Xi Jinping. The model advised users to use respectful language and redirected the conversation toward China’s development.

Hirundo claims to have developed a method they describe as “AI brain surgery” to remove these biases. By editing the model’s neural weights rather than relying on superficial rule-based prompts, the firm created a “Westernized” version of Qwen. According to Ben Luria, Hirundo’s CEO and founder, the original model exhibited censorship or propaganda-aligned bias in nearly 90% of sensitive political prompts, whereas the modified version did so in only 2.8% of cases.

Security experts have raised alarms about the adoption of Chinese AI models. A June report from Booz Allen Hamilton found that when Qwen was tasked with writing code for the U.S. government, the output contained 130% more security vulnerabilities than expected. Similarly, CrowdStrike reported that Chinese models exhibited trustworthiness issues in prior testing.

Alibaba did not respond to requests for comment regarding the findings. U.S. companies currently utilizing Qwen, including Airbnb and Uber, also declined to comment. However, usage data from OpenRouter shows that Chinese open-weight models have surged from less than 2% of global usage in late 2024 to over 45% by June 2026, underscoring the scale of the issue.