Saturday, Nov 21 | --:--
Back to home

House Chairmen Urge Ban on Chinese AI Models in Federal Systems and US Open-Weight Push

House Homeland Security Chairman Andrew Garbarino and China Select Committee Chairman John Moolenaar released a joint report on November 20 on US companies' use of Chinese open-weight AI models, using Airbnb, Anysphere and DoorDash as case studies. Its 12 recommendations include prohibiting PRC-developed models and their derivatives in federal information systems and federal contract work, barring them from high-risk critical-infrastructure functions, a provenance standard and AI bill of materials, and a national American open-weight AI initiative. The recommendations are not law.

5 min read Washington, D.C. View as Markdown
A handmade editorial collage on torn ivory paper in ultramarine and saffron: a large shallow wooden-sided paper seed tray holds rows of tiny sprouting paper seedlings with slender stems and two small sage or saffron leaves, while one section of the tray has been lifted out by its own wooden edge and sits beside it, leaving a gap in the tray; no hands, no flags, no logos.

The most consequential line in Congress's new report on Chinese AI is an admission about American AI: US companies are choosing Chinese open-weight models because nothing domestic competes at the price. The joint report released on November 20 by House Homeland Security Committee Chairman Andrew Garbarino and House Select Committee on China Chairman John Moolenaar pairs that finding with a call to bar PRC-developed models from federal systems, and its first recommendation is to build the alternative.

What the chairmen recommend#

The report makes 12 recommendations. The sharpest would prohibit PRC-developed AI models and their derivatives in federal information systems and in the performance of federal contracts, with a narrow waiver process for research, testing and defensive security work. It would also prohibit them in high-risk functions of critical infrastructure, with removal schedules tied to the availability of trusted alternatives.

Recommendation 1 is a national American open-weight AI initiative pairing public compute, shared evaluation infrastructure and lawful access to American frontier models for training with federal procurement demand. Recommendation 3 would have NIST, with CISA, develop a model provenance standard and an AI bill of materials, required for federal contractors and high-risk critical-infrastructure functions and voluntary elsewhere. Others call for sanctions and Entity List actions against firms engaged in industrial-scale illicit distillation, and for examining sales of American training data to designated PRC labs.

The three case studies#

The committees opened the investigation in April and examined three companies. Airbnb runs customized versions of Alibaba's open-weight Qwen models in customer support, including an assistant handling roughly 80,000 guest and host interactions a day, the report says. DoorDash reported that 25 PRC model families had been authorized and deployed in a non-trivial manner, and that replacing one embedded model could take up to six months. Anysphere built Composer 2, the flagship model in its Cursor coding platform, on a Moonshot AI base model, and told the committees it plans to explore phasing out PRC-developed open-weight models because of geopolitical risk, according to the report. Airbnb's use of Qwen was already at the centre of the administration's reported disclosure idea.

The findings behind them#

PRC models accounted for 41% of downloads on the leading open-model repository over the preceding year, the report says. In the most recent independent ranking of open-weight models the committees reviewed, the highest-ranked American model placed tenth, behind six PRC models. Companies examined cited the lack of competitive American open-weight options as a principal reason for their choices.

The report also says US agencies assessed that six leading PRC developers conducted industrial-scale distillation campaigns against American frontier models, likely with PRC government awareness, which tracks Reuters' account of distillation moving into Chinese industry. American data-labeling firms reportedly earn about $500 million a year selling training data to PRC labs, it adds, citing Forbes.

The Record reports that none of the three companies immediately responded to its request for comment. Mark Montgomery of the Foundation for Defense of Democracies told The Record the report matters because it explains the risks of the US not competing seriously in the low-cost open-model market.

Our read: a federal ban is the easy half. The report's own evidence says adoption is driven by cost and performance, so unless Recommendation 1 is funded, restrictions would move Chinese models out of government systems without changing what private companies choose.

Limits#

  • These are recommendations from two committee chairmen, both Republicans; nothing here is law.
  • Company usage details come from the companies' responses as summarised by the committees; we have not verified them independently.
  • The distillation assessment is attributed to US agencies via the report; the underlying classified briefings are not public.
  • The ranking finding refers to one independent index as of September; rankings change quickly.

Sources#

Prior Coverage

Earlier Times of AI reporting on this thread.

Scroll to continue reading