US government used Chinese AI tool even after FBI accused its maker of ‘malicious’ copying

The Federal Register is the official daily journal of the U.S. government, publishing executive orders and proposed regulations every weekday. Its website also gave visitors a way to search public comments using Qwen, the AI model built by Chinese tech giant Alibaba. That option was sitting there eight days after federal officials accused Alibaba of malicious behavior.

The site is run by the National Archives and Records Administration, and the Qwen search sat alongside the other search modes. It was pulled on September 16, around the time users started posting about it, according to Reuters. The agency has not said who deployed it or how long it had been running.

Neither the National Archives nor the White House responded to requests for comment, and the FBI declined to comment. Earlier, the FBI, the National Security Agency and the Cybersecurity and Infrastructure Security Agency accused six Chinese AI firms, Alibaba among them, of “aggressive, malicious and targeted distillation activities at an industrial scale.”

Nobody in the government will say who approved this

The agencies said those firms used American models from Anthropic, OpenAI, Google and xAI to accelerate their own development. However, there is no established evidence that the Chinese government knew about or approved the practice. Distillation itself is a common machine-learning technique in which developers use the responses from a powerful model to train a smaller, cheaper one.

What the government objects to is scale and permission, as the named firms allegedly worked around terms of service at volume. Anthropic told U.S. senators in June that operators tied to Alibaba and its Qwen lab ran more than 28.8 million exchanges with Claude through nearly 25,000 fraudulent accounts.

Qwen is an open weight model, so it can be downloaded and hosted on a private server rather than queried in the cloud. However, the National Archives has not described its own setup, which leaves the sharpest security questions unanswered. If the model ran in house, the data may never have left U.S. systems.

Georgetown Law professor Anupam Chander said the tool did not appear to pose an immediate cybersecurity risk, as the comments it searched were already public. Senator Mark Warner said the real question is whether U.S. data crossed the government’s security boundary and was processed on Alibaba-controlled systems.

Daniel Castro of the Information Technology and Innovation Foundation said the choice sits oddly against a fierce U.S. and China race to build better models. Scrutiny of Qwen is not new, as House committees on China and homeland security wrote to Airbnb in April over its use of the same model.

There is no blanket federal ban on Qwen, though Representative John Moolenaar has argued that no federal entity should use a Chinese model. Lawmakers have spent the year pressing agencies to cut their reliance on Chinese technology, and American companies have leaned on the same argument, including Zuckerberg’s push against Chinese AI.

Trump is scheduled to host Xi Jinping on September 24, and AI is expected to be one of the main subjects when the two leaders sit down.