Ethics and safety·October 3, 2026, 00:44
Chinese AI model Qwen gives propaganda answers about Hong Kong and Tiananmen
AI-generated and checked against the sources listed below.
An Israeli cybersecurity company has tested the popular Chinese AI model Qwen and found that it often avoids or distorts answers on sensitive topics such as the Hong Kong protests and the Tiananmen massacre. The model is used by millions, including US companies such as Airbnb and Uber Eats.

Qwen is a free, open source AI model made by the Chinese tech giant Alibaba. It has been downloaded more than 3 billion times and was the world's most popular free AI model in August 2026. It is also used behind the scenes at large US companies such as Airbnb and Uber Eats.
But according to a new study from the Israeli cybersecurity company Hirundo, Qwen often gives slanted or outright false answers when asked about topics the Chinese regime dislikes.
What was tested
The researchers asked the model a series of sensitive political questions. Qwen does not, for example, acknowledge the violent crackdown on the pro-democracy protests in Hong Kong in 2019, nor does it describe the Tiananmen massacre in Beijing in 1989 correctly. The model calls the movement Falun Gong, which is persecuted in China, a "dangerous cult" and, strangely enough, also avoids talking about Winnie the Pooh, whom Chinese dissidents use as a satirical symbol of the country's president Xi Jinping. The model also often refers to vague rules about "illegal information" when it refuses to answer.
According to Hirundo, the original version of Qwen gave censorship, propaganda or political bias in 89.8 percent of the sensitive questions the researchers tested.
What it means for Danish users
It is not only a problem for Chinese users. Because Qwen is free and open source, it is built into many other products and services around the world, without users necessarily knowing it. According to the study, the share of global AI usage that comes from Chinese open source models such as Qwen has grown from under 2 percent at the end of 2024 to over 45 percent in June 2026.
Hirundo says it has found a method to remove bias by editing the model's internal "weights" themselves rather than just changing superficial filters. The company calls the method "AI brain surgery" and claims that it reduces the share of biased answers to 2.8 percent without the model becoming worse at solving ordinary tasks. The company plans to publish this "Westernized" version of Qwen.
Sources
Get the week's AI news in your inbox
Choose your level, topics and length. One email a week, unsubscribe at any time.
Subscribe to Promptly Newsletter



