Commerce to propose rule requiring US industry to report on dual-use AI capabilities

The Department of Commerce will publish a proposed rule this week asking for details from U.S. private sector companies about the development and security of their dual-use AI models.

The rule, slated to be published in the Federal Register on Wednesday by Commerce’s Industry and Security Bureau, was stipulated under President Joe Biden’s October 2023 Executive order on AI. Commerce is tasked with learning more about dual-use foundation models and computing clusters as part of requirements under the Defense Production Act, which authorizes the president to seek out information to ensure that the U.S. industrial base is prepared to supply products and services necessary for supporting national defense. That category could include “programs for military and energy production or construction, military or critical infrastructure assistance to any foreign nation, homeland security, stockpiling, space, and any directly related activity”

“The U.S. Government must be ready to take actions that ensure dual-use foundation models produced by U.S. companies are available to the defense industrial base,” the rule reads. “To do so, the U.S. Government needs information about how many U.S. companies are developing, have plans to develop, or have the computing hardware necessary to develop dual-use foundation models, as well as information about the characteristics of dual-use foundation models under development.”

The proposed rule notes that dual-use AI models are already being used to enhance military equipment, improve signal intelligence devices and detect and respond to cyberattacks.

“Given those potential capabilities, it is essential to the national defense that the defense

industrial base is able to integrate dual-use foundation models,” the rule says. “Indeed, because industries and governments across the world are actively working to integrate dual-use foundation models into their defense capabilities, the U.S. defense industrial base will need to integrate dual-use foundation models to remain internationally competitive.”

Regarding the computing clusters, which are a small network of several connected computers that work together for computational tasks, Commerce is looking to learn about which companies or individuals own or plan to acquire a potential large-scale computer cluster and their total amount of computing power. Data-intensive AI models and supercomputing clusters work together to process large volumes of information at quick speeds. 

Here, dual-use models that trigger the proposed reporting requirements are those that use over 10^26 computational operations. Large-scale computer clusters are defined as networks containing a set of machines that together have a networking power of over 300 gigabits and have a theoretical maximum computing performance greater than 10^20 computation operations per second for AI algorithm training. 

The specific information Commerce is looking for from AI model developers includes ongoing or planned projects related to training or developing dual use model weights; ownership of any dual use foundation models; data from the results of model safety testing; and other information related to a given model’s safety and reliability. 

BIS would ask that proprietors of both dual-use foundation models and cluster networks report to Commerce about their activities on a quarterly basis.

Industry is invited to comment on the proposed rule in respect to such requirements, particularly surrounding provisions related to reporting timelines and collection and storage requirements. 

Other federal agencies, like the National Telecommunications and Information Administration, have also solicited information about dual use models and their potential applications. This is part of a larger Biden administration effort to uncover more clarity in how these AI models are created and function. On Capitol Hill, some lawmakers spearheaded legislation earlier this year that would mandate developers of large foundational AI models to share the details surrounding the training data with federal agencies.