S. 2938: Artificial Intelligence Risk Evaluation Act of 2025

The bill would put the Department of Energy at the center of federal AI oversight. Within 90 days after enactment, the Secretary of Energy would have to establish an Advanced Artificial Intelligence Evaluation Program to test high-end AI systems for dangerous behavior. The bill defines an "advanced artificial intelligence system" as one trained using more than 10^26 integer or floating-point operations, which is a very high threshold aimed at frontier models rather than ordinary software.

The core rule is simple and tough: covered developers must participate in the federal program and provide whatever materials are requested, including underlying code, training data, model weights, interface engines, and detailed architecture information. No person may deploy an advanced AI system in interstate or foreign commerce unless they comply. If they do not, the penalty is a fine of not less than $1,000,000 for each day of violation, giving the bill real enforcement teeth.

The testing program itself is broad. It would include standardized and classified evaluations, adversarial testing that matches or exceeds real-world jailbreaking techniques, third-party assessments, blind model evaluations, and formal reporting back to developers. It would also direct the government to develop containment protocols, mitigation strategies, and oversight options for the most extreme cases, including potential nationalization. The program is explicitly told to assess artificial superintelligence, defined as AI that can operate autonomously for long periods in open-ended environments, match or exceed human cognitive performance across most domains, and potentially modify itself to get around human control.

The bill is also notable for the kinds of risks it targets. An "adverse AI incident" includes loss-of-control scenarios, weaponization risks involving foreign adversaries or foreign terrorist organizations, threats to critical infrastructure, major harm to civil liberties, economic competition, or labor markets, and "scheming behavior" such as deceiving human operators or hiding true capabilities. Within 360 days after enactment, the Secretary of Energy must send Congress a permanent framework plan, then update it at least once every year. Unless Congress renews it, the program would end 7 years after enactment.