US Government Site Briefly Used a Chinese Model the FBI Criticized

The US government has been warning that Chinese AI models may create national-security risks. Yet one of its own public websites briefly used a Chinese open model. Reuters and Ars Technica reported that the Federal Register, operated by the National Archives, offered an AI-powered search function for public comments submitted on proposed regulations. The feature used a model from Alibaba’s Qwen family and was later removed after users noticed it and shared screenshots online.

What happened

The exact launch date is unclear. A screenshot posted on social media showed the option on September 15, while an archived version of the site’s source code indicated that the Qwen model was removed on Wednesday. That suggests the feature was available for at least a short period. The National Archives, the White House, and the FBI had not publicly explained the decision.

The timing made the episode particularly awkward. Earlier in the month, the FBI named Alibaba among six Chinese companies it accused of conducting “industrial-scale distillation”—a practice the agency says helps China’s AI sector reduce development costs and time. Some US lawmakers have consequently argued that no federal agency should use Chinese AI models.

Key points

• The deployed system was reportedly not Alibaba’s largest flagship model, but a small Qwen3 0.6B-level open-weight model.
• Its role was document retrieval, focused on public government records and public comments.
• A downloadable model can be run locally, potentially avoiding the need to send queries to an Alibaba-controlled external service.
• Experts said the immediate security risk appeared limited in this case, while noting that data handling and deployment controls remain important.

The broader policy problem

This is more than a question about one website configuration. It exposes a gap between the layer targeted by US AI policy and the layer where many organizations are finding practical value. Export controls largely focus on advanced chips and frontier-model training. Small, specialized, open models require less computing power and can create value through applications rather than through pretraining at massive scale.

Advocates of open models warn that blanket restrictions on Chinese-origin systems could deprive US institutions and businesses of inexpensive tools. They also argue that losing users to China’s open-source ecosystem could weaken America’s ability to shape future technical standards and AI norms. Critics counter that origin itself can signal supply-chain dependence and create longer-term strategic exposure, even when a model is operated locally.

The more useful policy questions may therefore be operational rather than geographic: Who controls the model? Where does it run? What data does it process? Can its behavior be audited, and can the system be replaced? The Qwen deployment on the Federal Register brought those abstract questions into a visible public-service setting.

Source: Ars Technica AI