Federal Register took down its Qwen search the same day

For at least part of Wednesday, you could search public comments on proposed US federal regulations with a tool built on Alibaba's Qwen model, on a website run by the National Archives, Reuters reports. The option disappeared later the same day, around the time social media posts about it started showing up.

At a glance

• The FBI accused Alibaba last week of maliciously copying Anthropic's models, and US officials said Chinese AI companies were copying American technology on an industrial scale.
• Qwen is open-weight: key components are published, so developers can download the model, modify it for a specific job and run it, often at lower cost than closed models.
• Georgetown Law professor Anupam Chander said he saw no immediate cybersecurity risk because the site's content is already public, though he added that it depends on how the model was trained.

If you missed last week, here is the setup. The FBI accused Alibaba of copying Anthropic's models to build its own AI tools, and US officials described the copying by Chinese AI companies as "malicious" and happening on an "industrial scale". A spokesperson for the Chinese embassy in Washington called the allegations unfounded. The accusations landed ahead of a meeting next week between the two countries' leaders.

The Qwen option sat on the government's daily list of proposed rules

The Federal Register publishes the daily list of newly proposed federal regulations, and the National Archives runs the site. Public comments are the raw material of that process: anyone can file one on a proposed rule, and on a contested rule the pile gets long enough that reading it end to end is not realistic.

On Wednesday, the site offered a Qwen-powered tool for browsing those comments alongside its other search options. Reuters could not establish when it was deployed. The tool came down on Wednesday, around the same time the social media posts appeared, according to screenshots and an archived version of the site's source code reviewed by Reuters.

Castro called it "one of the most insane things I've ever seen"

Daniel Castro, president of the Information Technology and Innovation Foundation, said there was a disconnect between the fierce US-China competition to build better AI models and a US agency choosing a Chinese-made one. That is where his line about the most insane thing he has seen comes from.

Representative John Moolenaar, the Michigan Republican who chairs the House China Committee, said no federal government entity should use a Chinese AI model, because doing so only makes the federal government more dependent on Chinese AI models and that is not in the national interest.

The committee has asked this kind of question before. In April, the House committees on China and homeland security wrote to Airbnb, which also uses Qwen, asking for details about that use as part of an investigation into the national security risks created by integrating Chinese AI tools into platforms Americans use.

Does a Qwen-powered search send anything to Alibaba?

Not necessarily, and that is what open weights change. Qwen's key components are published, so a developer can download the model, adapt it and run it on their own hardware. Think of it as buying the recipe instead of ordering from someone else's kitchen: the cooking happens where you already are.

Security experts say that is precisely the appeal, since running a model on your own IT infrastructure gives an organization more control over data that would otherwise go to an outside provider. Senator Mark Warner, the Virginia Democrat who is vice chairman of the Senate Intelligence Committee, said in a statement that the risk here depends on whether US data left the government's security boundary and was processed on Alibaba-controlled systems.

Chander's read was similar in shape: the use of Qwen did not appear to present an immediate cybersecurity risk, though that depends on how the model was trained, and the Federal Register's content is public to begin with.

The missing pieces are the ones that would settle the argument: who built the search, where the model actually ran, and when it went live. AI experts said the tool may not have posed national security risks, and the question they raised was about the federal government contradicting its own messaging on Chinese AI models rather than about leaked data. Oddly, on a site whose entire job is publishing the government's paperwork in the open, the clearest surviving trace of the feature is an archived copy of the source code.

Before next week's leaders' meeting

Three things are worth watching. Whether the National Archives says what the search was and where it ran. Whether the Qwen option comes back in any form, or is replaced by something built on a US model. And whether the House China Committee's Airbnb inquiry, opened in April, produces anything public about how Qwen gets embedded in services Americans use. The leaders meet next week.