Regulatory intelligence for US exporters

Commerce Hears Arguments Favoring and Against Export Controls on AI Models

The Commerce Department should start preparing export controls for dual-use artificial intelligence models, which could prevent those models from being used to make biosecurity weapons or skirt U.S. export restrictions on advanced semiconductors, researchers told the agency in comments released this month. But technology companies and industry groups warned the U.S. against overbroad controls, which they said could hurt American AI innovation.

Start A Trial

Commerce in February requested public comments on the risks, benefits and potential policy actions it should take to address advanced AI models with widely available model weights, which could make AI tools more available to small companies, researchers and nonprofits (see 2402220025). The agency’s National Telecommunications and Information Administration, which was seeking feedback as part of President Joe Biden’s October AI executive order (see 2310300029), said widely available model weights could help more people benefit from AI but also raise the risk of foreign adversaries using AI in ways that harm U.S. national security.

U.S. technology industry officials said Commerce needs to strike the right balance with any new AI export controls.

“Imposing the highest tier of government oversight on all large-scale foundation models would essentially subject all large foundation models to ongoing government requirements and oversight requirements” under the executive order, along with “export control requirements overseen by” the Bureau of Industry and Security, the Software & Information Industry Association said.

Those export controls could “limit AI innovation, research, and real-world applications without leading to measurable improvements in risk mitigation,” the association said. “We do not think this was the intent of the EO.”

Microsoft made similar points, saying AI developers should be allowed to “responsibly release AI models under their preferred licensing terms and distribution method,” which “is essential to innovate, compete, and partner in the global marketplace.” U.S. rules “that put undue burden on these decisions can have chilling and unintended consequences on America’s innovation and competitiveness,” the company said.

It urged NTIA to work with BIS to make sure its definitions for “open source” or “widely available” foundation models align with the “published” definition under the Export Administration Regulations. Under the EAR, Microsoft said, a model “must be freely and publicly available without limitation on distribution” to qualify as “published.”

“It is crucial that the ‘open source’ or ‘widely available’ terms used by the Federal government are in harmony with the EAR’s ‘published’ standards to facilitate the unencumbered distribution of these models, thereby avoiding unintended export control implications,” the company said.

Other commenters advocated for new export controls, including the Center for a New American Security, which said BIS should develop a “comprehensive, integrated strategy” to address AI export risks. That should include AI technologies “all the way from semiconductor manufacture through to model weights themselves,” CNAS said.

The think tank said existing U.S. chip export controls are limiting China’s ability to train “compute-intensive foundation models,” but widely available model weights “effectively circumvent these controls, allowing Chinese AI labs to download models which they themselves may not be able to train -- or for which training may be cost-prohibitive -- given these controls.”

CNAS acknowledged that it’s challenging to decide whether and how far to pursue export controls on AI models and architectures, partly because the U.S. may want to keep its adversaries reliant on American models rather than incentivize them to create their own.

“As with U.S. chip exports, policymakers must weigh when it makes sense to foster this dependency, when cutting adversaries off may be wise, and to what extent doing so may accelerate indigenous capabilities,” CNAS said. “[A] potentially attractive strategy would be encouraging the diffusion of models and architectures near or slightly ahead of Chinese equivalents, while discouraging this for the most advanced U.S. capabilities.”

Johns Hopkins University also said the U.S. should start preparing possible export controls, but they should be “narrowly targeted” toward dual-use foundation models with “concerning biosecurity capabilities.” The school’s Center for Health Security specifically said the U.S. should start working with members of the multilateral Australia Group to discuss how they could update export controls to “reduce high-consequence biosecurity risks associated” with AI models.

NTIA should think about what “capabilities” would justify export controls, how those restrictions “might be narrowly tailored to apply only to the most concerning set of dual-use foundation models,” and whether other models posing similar risks may be “readily available from sources” the U.S. and its allies can’t control, the school said.

It added that some open models can “confer significant social benefits, including to public health” but stressed that they also can be used to produce biosecurity weapons. “For these reasons, any export controls on open dual-use foundation model weights should be narrowly tailored to address high-consequence threats to safety,” Johns Hopkins said, “such as the high-consequence biosecurity risks.”

Still, the university said it isn’t suggesting Commerce place export controls on dual-use foundation models “today.” But “the risks posed by open, biologically capable dual-use foundation models are grave enough for the US government to prepare such policy options so they can be deployed when and if they become relevant.”

Johns Hopkins and other commenters also warned about the possible “legal challenges” the U.S. may face in “applying export controls in this domain.” A professor with Cornell Law School’s First Amendment Clinic pointed to Bernstein v. U.S., a more than two-decades-old case before the U.S. Court of Appeals for the 9th Circuit, which ruled that export restrictions on computer source code are a restriction on speech and violated the First Amendment.

Although the 9th Circuit later withdrew that opinion, Cornell professor G.S. Hans said other courts have “relied upon the underlying reasoning,” and AI companies “may rely upon the reasoning of Bernstein to argue that export restrictions on model weights violate” the Constitution.

“[A]ny such export control regulations would also almost certainly face First Amendment challenge,” Hans said. “If other courts follow the reasoning of Bernstein, such challenges present a serious obstacle to export control regulations.”