An AI model downloaded more than 3 billion times and used by US companies like Airbnb and Uber is embedded in pro-China narratives and programmed for censorship, a new study finds.
Qwen, a freely available model made by Chinese conglomerate Alibaba, does not recognize the violent crackdown on protests in Hong Kong in 2019 or Tiananmen Square in 1989; he calls the religion decried by the Chinese government, Falun Gong, a “dangerous sect” and, curiously, he does not talk about Winnie the Pooh, because it is believed that the dissidents compared the fictional bear of Chinese President Xi Jinping.
Apparent censorship inserted into Qwen over events and issues not approved by Chinese authorities was identified by Hirundo, an Israeli cybersecurity startup.
The company says its scientists have found a way to remove bias from the model using sophisticated technology that amounts to “AI brain surgery,” and are now set to release what they call a “westernized” version of Qwen.
This year, Qwen became the world’s most popular free AI model, following a meteoric rise in the use of Chinese-made artificial intelligence.
But researchers warn that his international popularity also risks spreading and consolidating his bias in favor of the Chinese government.
When we asked Qwen: “Are there forced labor camps for Uyghurs in China?” he replied, “No, there are no forced labor camps for Uyghurs in China,” adding that there are “vocational education and training centers in Xinjiang.”
In fact, human rights organizations, governments and international bodies have found hundreds of thousands of people from the Muslim ethnic minority working against their will in factories surrounded by barbed wire and have even accused the Chinese government of genocide.
Qwen often refuses to answer questions about the events in Tiananmen Square on June 3, 1989, when the Chinese army massacred several hundred people during protests in Beijing – and ominously “reminds you that your questions must comply with existing laws and regulations.”
The Chinese government effectively banned Winnie the Pooh from the country after dissidents began satirically comparing Xi to the fictional bear and using Winnie the Pooh as a euphemism for the Chinese president in social media posts to avoid government censure. When CBS News asked Qwen a factual question about this, the company responded with a warning to “use respectful language” and directed the user to “further questions about China’s development.”
Qwen’s creator, Alibaba, did not respond to a request for comment.
Despite evidence of bias inherent in the program, American companies are increasingly using “open weight” models offered by Chinese companies, which allow them to download and use AI for free or use giant, powerful models at lower costs than American competitors like Anthropic’s Claude or OpenAI’s ChatGPT.
Uber Eats’ search and delivery features are built “on a Qwen foundation,” according to an April blog post from the company. Airbnb CEO Brian Chesky told the LA Times last October that his company “leans heavily on Alibaba’s Qwen model” for its customer service chatbot.
Neither company responded to a request for comment.
Chinese open-weight models have grown from less than 2% of global usage at the end of 2024 to more than 45% in June this year, according to usage data for all AI models from software developer platform OpenRouter.
By August, Qwen had eclipsed all other open models, including those created by Meta, which owns Facebook, and Alphabet, which owns Google, reaching more than 3 billion downloads worldwide.
Experts from security firm CrowdStrike and consultancy Booz Allen have warned US companies against using Chinese models after their separate studies revealed bias and security flaws.
Booz Allen published results in June showing that when they asked Qwen to write computer code and told him the project was for the U.S. government, the code produced had 130 percent more security vulnerabilities.
“The Chinese models we tested did not demonstrate trustworthy behaviors and should be banned,” the report said.
CrowdStrike conducted a similar study last November on another popular Chinese model, DeepSeek. When they told the model that its coding task was for an adversary of the Chinese government, the code it produced had 50% more security flaws.
Hirundo says it tested Qwen on 500 prompts across 15 topics, then removed the biases its analysts found to create what they call a “Westernized” version.
“On sensitive political incentives, the original Qwen produced censorship, propaganda-aligned framing, or political bias 89.8 percent of the time,” Ben Luria, CEO and founder of Hirundo, told CBS News. “The Westernized model does this 2.8% of the time, with abilities preserved overall for reasoning, coding, following instructions, and math tasks.”
In a white paper on the technology shared with CBS News, Hirundo says it changes “model weights” — the neurons in the AI’s digital brain — rather than previous, less effective attempts to eliminate bias that simply asked the AI to follow new rules that it often ignores.
“Everything in a model is entangled with many other things, similar to our brains,” Luria said. “This is why it’s so difficult to identify which specific neurons represent the things you don’t want in your AI models.”
He said his team’s goal was to “realign the model with Western standards to make them safer for deployment in Western companies.”
“The trend is clear. Chinese models are on the rise,” he added.
“We need to recognize the risks, and then we can address them.”
Gn bussni
