House committee chairmen urge removal of Chinese AI models from federal systems
Two US House of Representatives committee chairmen recommend banning Chinese AI models from federal systems and contract work, and the ban would cover more than DeepSeek. None of the report’s 12 recommendations has become law.
Republicans Andrew Garbarino and John Moolenaar released the report on Friday. Garbarino chairs the House Homeland Security Committee and Moolenaar the House Select Committee on China. Existing federal restrictions cover DeepSeek, whose technology has been barred from federal civilian agency systems by a Department of Homeland Security order, Turkish news agency Anadolu Ajansı reported. The recommendations say Chinese models should also be removed from sensitive tasks in critical infrastructure. The committees plan to continue monitoring the issue.
The concern centres on open-weight models. Their model weights, the parameters learned during training, can be downloaded and run on an organisation’s own servers. The report says companies are adopting Chinese open-weight models because they are competitive and much cheaper. It also notes that the US has no domestic open-weight models that match them on price and performance.
India’s Lokmat Times published a story from the IANS news agency reporting that Airbnb runs customised versions of Alibaba’s Qwen models on US-based servers. The models handle about 80,000 daily exchanges and summarise about 250,000 customer support cases. DoorDash uses 25 Chinese model families, and replacing one key model would take an estimated six months. Lokmat says it edited the story with AI and that a journalist has not checked it. The companies say they have no commercial agreements with Chinese developers. The committees nonetheless warn of growing dependence. The report did not show that the models the companies use had led to a cyberattack or data breach, Anadolu reported.
Users may not be able to see which model runs behind a product. Anysphere, which develops the Cursor coding tool, built its Composer 2 model on a base model from China’s Moonshot AI. The report proposes an AI bill of materials, which would show buyers the models and components in a system.
The committees also reviewed procurement documents from China’s People’s Liberation Army, India’s RTV reported. In a notice dated 10 July, the army requested DeepSeek models for a cyber training centre, for tasks described as ‘intelligent attack’ and ‘intelligent intrusion’. The documents show intentions. They do not show that the models have been used in real cyber operations, and the report’s claims have not been independently verified. Under the army’s plans, the models would run on closed networks without internet access. Open-weight models are easier to deploy on such isolated networks.
The Wisevoter website says the restrictions could extend to more model families. It therefore advises organisations that use open-weight models to record where each model runs and who has access to it.
Read more: Models