Yesterday, three Republican and three Democratic legislators introduced new federal AI regulation in the U.S. House of Representatives, termed the FRONTIER Act. The new bill was adapted from portions of the Great American AI Act (GAAIA) draft, which StopWatch covered in June and mentioned in yesterday’s roundup of policy proposals.
The FRONTIER Act’s lead sponsors, Congressman Jay Obernolte (R-CA) and Congresswoman Lori Trahan (D-MA), also led the drafting of GAAIA, so their collaboration here makes sense. The bill made some significant changes, though, and some of its provisions are the first of their kind.
Among other things, the FRONTIER Act would require safety frameworks, third-party audits, independent verification, and incident reporting. In principle, this means that large AI companies would still get to write their own standards, but the government (with third-party help) would make sure they are followed. This is supported by a mandate to report “critical safety incidents” within days.
These requirements resemble existing state regulations like California’s SB 53 and New York’s RAISE Act, but there’s more. Licensed third party organizations would also review AI companies’ frameworks to determine if they are adequate. (Which, for the moment, they clearly are not.)
The bill would appoint a new Under Secretary for AI Security (within the Department of Commerce) to administer the federal rules. It would also empower the Secretary of Commerce to issue an emergency order
suspending or restricting a frontier developer’s development, deployment, or internal use of a frontier model upon finding that such development, deployment, or internal use of that model presents an imminent catastrophic risk.
The FRONTIER Act’s requirements would be enforced by a fine of up to $1M per day of noncompliance, or $10M a day for emergency orders. Willfully ignoring an emergency order would be a crime punishable by years of jail time. This is a pretty big deal; as far as I know, no other AI safety law carries criminal penalties.
Despite these new provisions, the bill does contain some concerning loopholes and omissions. Independent verifiers would need a federal license, but nothing obliges the new Under Secretary to issue one, and the requirement doesn’t kick in until a year after they do. If Commerce drags its feet, or struggles to find competent verifiers, we could end up with nothing for a very long time.
As Sam Sabin of Axios reminded us just today, current AI models are so capable and cheat so hard on tests that third-party evaluators can barely keep up. Risk assessors face a somewhat different challenge, but might find themselves similarly overwhelmed.
The U.S. Center for AI Standards and Innovation (CAISI), a federal agency tasked with understanding and evaluating frontier AI, might be able to help pick up the slack. But the proposed $100M a year in funding for CAISI that was present in the GAAIA draft is missing from the current FRONTIER Act, and I’m not sure why.
Also absent are the whistleblower protections that GAAIA originally proposed. I’m disappointed to see that, because whistleblower protections seem like one of the easiest and most straightforwardly helpful policies the U.S. could implement.
Another potential concern: The FRONTIER Act’s “emergency order” for shutdown mirrors the intent of the also-bipartisan AI Kill Switch Act, but it also covers internal use of a model, not just public deployment. But the Kill Switch Act has other advantages: it explicitly requires companies to build in the capability to pull access, and it can be triggered proactively by incidents of AI misbehavior, not just costly or lethal disasters.
Right now the two bills conflict instead of working together, because the FRONTIER Act states that its emergency-order provision is exclusive. This means any other U.S. action to restrict AI model development or deployment by frontier companies would be illegal by default.
Honestly, I can imagine why this provision made it into the bill. Recent ad-hoc actions by the administration have no doubt made an impression on lawmakers, and I’m actually somewhat impressed with the bill authors for standing up for Congressional authority here.
But the FRONTIER Act also preempts any state laws “that imposes new substantive obligations on artificial intelligence developers with respect to any Covered Subject Area.” That includes transparency, auditing and verification, and incident reporting. This clause closes a lot of doors if the federal regime stalls due to verifier-licensing or other issues.
A clear, legible federal framework is better than a state-by-state patchwork in many ways, but only if that federal framework actually keeps us safe. The FRONTIER Act looks like a step in the right direction, but still a far cry from an international solution to the AI race.
The analyses and opinions expressed on AI StopWatch reflect the views of the individual contributors and the sources they cover, and should not be taken as official positions of the Machine Intelligence Research Institute.