What Hirundo’s Qwen tests and CBS’s checks actually show

Alibaba’s Qwen is an open-weight AI model used in products and services, including at Western companies such as Airbnb and Uber, according to CBS News. That makes questions about its answers more than a matter of chatbot accuracy: a model’s omissions or framing can carry into the products built around it. Hirundo’s findings and CBS’s checks point to possible political bias, but they do not establish that every Qwen version behaves alike—or that using Qwen exposes company data to Alibaba or the Chinese government.10

What Hirundo tested—and what its results mean

Hirundo, a cybersecurity startup, says it tested Qwen with 500 prompts spanning 15 topics. The company reported that on politically sensitive questions, Qwen produced censorship, propaganda-aligned framing, or political bias 89.8% of the time, according to a Hirundo executive quoted by Newsmax.11

That figure is Hirundo’s reported result, not an independently reproduced rate for all Qwen versions, prompts, or deployments. It also combines several kinds of response—evasion, aligned framing, and bias—rather than measuring refusal alone.11

Reported examples included questions about the 1989 Tiananmen Square killings and the 2019 Hong Kong protests. Hirundo’s results, as described in the reporting, included answers that would not confirm or acknowledge events; CBS’s coverage also highlighted Qwen’s responses about Hong Kong and other controversial topics.1115

What CBS’s checks add

CBS News separately questioned Qwen and reported alarming answers on sensitive subjects. Those checks offer an additional example of the behavior Hirundo described, but they are a spot-check, not evidence that every Qwen model or business deployment will respond in the same way.1015

The distinction matters: a model can avoid a topic, repeat a politically aligned framing, or provide an inaccurate answer without issuing an explicit refusal. Evaluating only whether it says “I can’t answer” could therefore miss other forms of omission or bias.11

Why businesses should care—and what the tests don’t prove

CBS reported that Qwen had been downloaded more than 3 billion times and was used by companies including Airbnb and Uber.10 If a business incorporates a model into customer-facing features or internal tools, its responses can influence what users see. That is a reason to test the specific model version and deployment for the tasks a company plans to use—not proof that a reported political-bias finding affects every product that uses Qwen.

Political-bias questions and security questions should also be kept separate. U.S. House committees have opened an investigation into national-security and cybersecurity risks associated with Chinese-developed AI models, including Airbnb’s reported use of Qwen. An investigation and the concerns it raises are not evidence that Qwen has caused a data breach.

Open-weight models can be run on a company’s own infrastructure, which may give the organization more control over sensitive data than sending it to an outside provider. That deployment choice does not, by itself, remove bias embedded in a model’s responses, and the details of any particular company’s setup matter.

Hirundo’s proposed fix remains a claim to evaluate

Hirundo says it has found a way to remove Qwen’s bias by editing the model’s internal weights, and CBS reported the company’s plan to release a “Westernized” version.10 That is a claim about a proposed technical remedy, not independent evidence that the change works reliably across topics, model versions, languages, or business tasks.

For organizations considering Qwen, the practical takeaway is to assess the exact model and deployment they intend to use. Test sensitive and routine prompts, check whether answers omit or distort relevant facts, and separately review data flows and security controls. The reported findings warrant scrutiny, but they do not settle either question for every use of the model.