Skip to content
Tech News
← Back to articles

US Federal Register Caught Using Chinese AI Model for Document Search

read original more articles
Why This Matters

The discovery that the US Federal Register's search tool briefly relied on Alibaba's Qwen3 model highlights a growing paradox: even government agencies are turning to cheap Chinese open-source AI despite political rhetoric warning against Chinese technology. This underscores the real-world cost pressures pushing US institutions toward foreign AI, raising questions about domestic AI competitiveness and national security messaging.

Key Takeaways

Sign up to see the future, today Can’t-miss innovations from the bleeding edge of science and tech Email address Sign Up Thank you!

Inexpenive AI models developed by Chinese companies are taking the United States by storm, as cost-conscious business leaders turn away from far more costly domestic AI.

From a business perspective, the incentive to do is clear — it’s all about the bottom-line at the end of the day. China’s open source models are much cheaper to run compared to the ones being developed by frontier labs based in the US, so it follows that businesses would choose the lower-cost option.

But as the capabilities of China’s homegrown AI models rise to match those of US models like Claude and ChatGPT, it seems the federal government has also taken notice.

A Switzerland-based commodity portfolio manager shared a screenshot of the document search interface of the Federal Register, the official journal of the US federal government. Far down on the site’s interface is an option to choose a “search mode,” which defaults to “semantic,” but also contains two options for “Qwen3:0.6B,” a large language model developed by the Chinese firm Alibaba Cloud.

“Even the United States government is unable to afford the cost per token charged by American AI companies,” one user replied.

A little over a day after that post began making the rounds, both Qwen options had been removed from the Federal Register search interface. Luckily, somebody thought to save a snapshot of the page on September 16, which preserves the two Qwen options.

Of the two flavors of Qwen-bot on offer, one is labeled a “hybrid” model, described as “512D, Recursive Splitting, Distilled, Prefixed,” while the other refers to Qweb3:0.6B, a relatively lightweight, half-a-billion-parameter model.

It’s an extremely ironic development as America’s top tech billionaires spout fearful rhetoric about the horrors of Chinese AI — a signal from the official journal of the US federal government that open-source Chinese models are good enough to get the job done.

The US government maintains the country is in a tight AI race with China, yet here it is using the country’s tech to make public records searchable. Just this week, Trump refused to pull the brakes on AI development — despite tech leaders increasingly calling for a slowdown — arguing that the US can’t afford to let China’s AI industry get ahead.

... continue reading