Bipartisan Senate bill would establish an AI safety office in Commerce

Five lawmakers from the Senate Commerce Committee introduced legislation to establish national security standards for frontier developers.

By Caroline Nihill

December 20, 2024

Listen to this article

0:00

This feature uses an automated voice, which may result in occasional errors in pronunciation, tone, or sentiment.

The U.S. House of Representatives and the U.S. Capitol Dome is seen as Speaker of the House Nancy Pelosi (D-CA) prepares to speak during a press conference on September 21, 2022 in Washington, DC. Speaker Pelosi, along with other lawmakers, is celebrating the passage of the Inflation Reduction Act and the steps it takes to lower healthcare costs. (Photo by Samuel Corum/Getty Images)

A bipartisan group of Senate lawmakers on Thursday introduced legislation that would establish an AI safety review office in the Department of Commerce, a move aimed at assisting frontier AI model developers.

The Preserving American Dominance in AI Act of 2024 (S.5616), sponsored by Sen. Mitt Romney, R-Utah., would require the office, led by a Senate-confirmed under secretary of Commerce, to develop and issue a take-home test for model developers to evaluate their systems for “extreme risks” before development. The bill would also require the newly established office to collaborate with the National Institute of Standards and Technology’s AI Safety Institute to provide best practices and technical assistance to industry, according to the release.

The bill — co-sponsored by Sens. Jerry Moran, R-Kan., Angus King, I-Maine., Jack Reed, D-R.I., and Maggie Hassan, D-N.H. — would also set reporting requirements and guidelines for frontier developers to prioritize national security implications of AI. That approach, the release states, is part of an effort to balance the need for competition against foreign adversaries.

Romney said in the release that he and his colleagues on the Senate Commerce, Science and Transportation Committee “have spent the last several months turning our framework into bipartisan legislation that would preserve America’s competitive edge in AI while also ensuring the most advanced models do not pose unchecked chemical, biological, radiological, nuclear, or cyber risks.”

Said Hassan: “We must continue to work to protect against potential risks to public safety, such as the misuse of this technology by terrorists or criminal organizations. By implementing balanced oversight of the most advanced AI models, we can maintain our technological leadership while protecting public safety and our national interests.”

Additionally, the office would study future AI-related risks and submit a report on the findings to Congress.

Frontier AI model developers, meanwhile, would be offered assistance from the Commerce office to identify best practices for red-teaming against chemical, biological, radiological, nuclear or cyber risks.

Under the legislation, developers would have to provide the AI safety review office a 90-day window to review a model’s safeguards against risks, ensure a “presumption of approval” to prevent unnecessary roadblocks to the deployment and implement cybersecurity standards.

Developers that do not comply with the legislation, should it become law, would face criminal and civil penalties. Criminal charges could result in imprisonment of up to 10 years. Civil charges would include a fine of a maximum of $1 million per day.

The bill also calls on Infrastructure-as-a-Service providers to implement “know-your-customer” standards for transactions with foreign individuals. And the most advanced data centers that are training the most advanced AI models would have to report the owner and facility location to Commerce’s AI office.