The Federal Register, the US government's official journal of record, has quietly stripped Qwen-powered search options from its advanced document search page, days after Reuters reported that the site had been offering the public a search tool built on a Chinese open-source model — a model family the FBI has said incorporates technology copied from Anthropic.
The removal went live without any public announcement or changelog. The episode is a small technical change with outsized symbolism: it shows how quickly Chinese-built AI can slip into American government systems through ordinary engineering decisions, and how fast it gets pulled once anyone looks. For more context on stories like this, see our ongoing AI policy coverage.
Five Search Modes Became Two
Until recently, the "Search Mode" section on the Federal Register's advanced search page offered users five ways to search the journal, according to a report by OfficeChai. There was a default Semantic mode, a Keyword Only mode, and a Neural mode built on msmarco-distilbert-base, a retrieval model from Microsoft's MS MARCO research. Alongside them sat two separate options built on Qwen3:0.6B, a small distilled model from Alibaba's open-weight Qwen family — one of them described as a hybrid retrieval setup with "512D, Recursive Splitting, Distilled, Prefixed" processing.
That is now gone. The page currently shows only Keyword Only and Semantic, with Semantic set as the default. Both the Neural mode and the two Qwen-based modes have been removed entirely, with no public changelog or announcement explaining the decision.
A Small Dropdown With Outsized Symbolism
The Federal Register is the government's official journal for rules, proposed rules, notices, and presidential documents, and it is run by the National Archives and Records Administration. A search option buried in an advanced-search dropdown is a minor technical detail on its own. It became notable because of where it sat: inside a federal system, at a moment when Washington has been working aggressively to keep Chinese AI out of the government's own stack.
As OfficeChai's report noted, that effort has ranged from the Commerce Department's ban on DeepSeek to legislation aimed at keeping any Chinese-linked model off government devices. Against that backdrop, the discovery that the government's official journal was letting the public run searches through Alibaba's model — and the quiet removal that followed — reads as a credibility problem rather than a routine infrastructure note.
The FBI's Copying Claim Hung Over the Story
Reuters' reporting on the episode carried an extra edge: the news agency described the Qwen search tool as one that the FBI has said copied Anthropic's technology. The claim stems from the ongoing controversy over whether Chinese labs have distilled outputs from American frontier models — an allegation that has shadowed the Qwen family as its open-weight releases have proliferated across Western infrastructure.
None of that changes what the Federal Register's engineering team actually deployed: a 0.6-billion-parameter model released under the permissive Apache 2.0 license, small enough to run as a search-ranking component on a public website. But the licensing says nothing about optics, and optics are what transformed a dropdown option into a story that spread across X and eventually reached wire services.
No Official Explanation — Yet
There is no confirmation of what prompted the removal after the Qwen option went viral on X. As OfficeChai pointed out, it could just as easily be routine search-infrastructure maintenance: the Federal Register's search tools have changed configurations before without explanation, and swapping out an embedding model for a search feature is a fairly unremarkable engineering decision in isolation.
But the timing — the modes vanished shortly after the Qwen setup was publicly reported — makes a security- or optics-driven rollback the more obvious read. Neither the National Archives and Records Administration nor Alibaba has issued a public statement explaining the change.
Open Weights Blur the Line
The deeper issue the episode exposes is that the line between "American" and "Chinese" AI inside government systems is far more porous than policy debates tend to assume. Qwen3:0.6B carries an Apache 2.0 license, meaning anyone — including a US government website — can download, modify, and deploy it without restriction. Open-source models do not carry passports, and engineering teams evaluating search components tend to pick them on benchmark scores and hosting cost, not country of origin.
For policymakers, the Federal Register episode is a preview of a recurring dilemma. Restricting Chinese-built closed APIs is administratively straightforward. Restricting open-weight models whose weights are mirrored across every major hosting platform is a different problem entirely — one that procurement rules, device bans, and commerce restrictions have not yet solved.
What Happens Next
The practical takeaway for federal agencies is that search infrastructure is now part of the AI policy perimeter. A dropdown that nobody flagged at deployment became a national story within days, and the default response — quiet removal — did little to clarify what standards, if any, governed the original decision.
Whether Congress or the National Archives responds with an actual review of how AI components get vetted in public-facing government systems remains to be seen. What the episode already demonstrates is that the vetting question is no longer hypothetical: the tools are small, the licenses are permissive, and the scrutiny arrives after the fact.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
