Proposed AI Reporting Requirements Tracking the Development of Foundational AI Models: Key Takeaways for Companies

This article was also published by the Harvard Law School Forum on Corporate Governance.

The Commerce Department’s Bureau of Industry and Security (BIS) has proposed a rule that would establish reporting requirements to track development of advanced artificial intelligence (AI) models, in accordance with instructions in an October 2023 executive order.

The proposed rule is intended to bolster the government’s understanding of the capabilities and security of dual-use foundational AI models.   BIS emphasizes that the U.S. government must be ready to take action to ensure the models are appropriately trained, can operate in a safe and reliable manner, and are not vulnerable to cyberattacks such that they are available to support the U.S. defense industrial base.  

BIS is seeking comments on the proposed rule, with all comments due by October 11, 2024.  BIS has expressed a particular interest in comments on the quarterly notification schedule, best practices for collecting and storing data submitted pursuant to the reporting requirements, and the technical thresholds that trigger the proposed reporting requirements.

The proposed regulation and underlying executive order demonstrate the seriousness of both (i) the U.S. government’s concerns about misuse of AI models in ways that would, in its view, undermine U.S. security, and (ii) the government’s belief that it is critical to have advanced AI models for use by the U.S. government.

Who Would the Rule Affect?

The proposed rule would require U.S. companies that are developing or have plans to develop dual-use foundational models to file quarterly reports with BIS.  The proposed rule would apply to U.S. persons that conduct or plan to conduct an AI training run utilizing more than 10²⁶ computational operations (e.g., integer or floating-point operations).

It would also require U.S. persons who acquire, develop, or possess large-scale computing clusters – defined as clusters having a set of machines transitively connected by networking of over 300 Gbit/s and having a theoretical maximum performance greater than 10²⁰ computational operations per second for AI training, without sparsity – to file such reports.

What Must be Reported?

BIS proposes that companies be required to submit a notification that they engage or plan to engage in covered activities.  Upon receipt of such a notification, BIS would send the submitting company a questionnaire for which a response is compulsory. 

Covered companies that develop dual-use foundational models would have to report:

• Ongoing or planned activities related to training, developing, or producing foundational models, including physical and cyber security protections that ensure the integrity of the training process;
• Ownership and possession of the model weights of any dual-use foundational models, as well as physical and cybersecurity protections for the model weights;
• Testing results for developed dual-use foundational models including results from reliability tests and red team testing, the presence of software vulnerabilities and exploits, the possibility of a model self-replicating or self-propagating, and the ability of the models to influence real or virtual events.  Companies would also have to disclose safety measures taken to improve performance on these tests and to improve the model’s overall security;
• Other relevant information that pertains to dual-use foundation models’ safety and reliability, as well as any related activities or risks that pose a potential threat to U.S. national security; and
• Information pertaining to acquisitions, development, or possessions of potential large-scale computing clusters, the location of the large-scale computing clusters, and each cluster’s total computing power.  

The proposal would require that responses to the questionnaire are submitted on a quarterly basis and cover activities that occurred during the preceding quarter or that are planned to occur in the six months following the quarter.