Energy Department would test advanced AI systems for risks

Covered developers would have to join the program and provide requested materials needed for its work. This applies to developers of advanced AI systems trained with more than 10^26 computing operations for use in interstate or foreign commerce. Requested materials may include code, training data, model settings, and details about how the system was built or trained.

A person could not deploy a covered advanced AI system for commerce without meeting the participation and information-sharing rules. The restriction covers deployment of advanced AI systems in interstate or foreign commerce. A violation would carry a fine of at least $1 million for each day it continues.

The Department of Energy would establish and run a program to test advanced AI systems and report on risks and safety measures. The program would evaluate advanced AI systems and provide reports to participating entities; it would also develop risk-reduction plans and possible oversight approaches. The program would use classified testing, including attempts to break the systems, and independent reviews when feasible. The Secretary would have to establish the program within 90 days after enactment. It would end seven years after enactment unless Congress renews it.

The Secretary of Energy would send Congress an oversight recommendation and update it each year while the program lasts. The recommendation would address federal oversight of advanced AI systems and risks identified through testing. The recommendation would draw on program data and include possible standards, licensing rules, monitoring, and government oversight options. The first recommendation would be due within 360 days after enactment. Updates would be due at least once a year for the program's duration.