S. 2938: Artificial Intelligence Risk Evaluation Act of 2025

S. 2938, the Artificial Intelligence Risk Evaluation Act of 2025, puts the Department of Energy in charge of stress-testing the most powerful AI systems before they reach the public. The Secretary of Energy would have 90 days after the bill becomes law to stand up an Advanced Artificial Intelligence Evaluation Program.

The program only targets the frontier. The bill defines a covered "advanced" system as one trained with more than 10^26 computing operations — a threshold aimed at the largest models, not everyday apps or smaller open models. The Secretary could later propose a new threshold, but it can't take effect unless Congress passes a joint resolution approving it.

The core obligation is blunt. If you build a covered system, you have to participate in the program and turn over what the government asks for: the underlying code, the training data, the model weights, the interface engine, and detailed architecture information. And you can't deploy the model until you comply. "Deploy" is defined broadly — selling it, providing API access, or releasing it open-source all count.

The testing itself goes well beyond a checklist. It includes classified evaluations, blind model tests, independent third-party assessments, and red-teaming meant to match or exceed real-world jailbreaking. The program is specifically directed to assess whether a system could reach "artificial superintelligence" — which the bill defines as AI that can run autonomously for long stretches, match or beat humans across most tasks, and potentially rewrite its own code to slip human control.

The risks the program hunts for are spelled out as "adverse AI incidents": loss-of-control scenarios, weaponization by foreign adversaries or terrorist groups, threats to critical infrastructure, erosion of civil liberties and labor markets, and "scheming behavior" where a model deceives its operators or hides what it can do. Within 360 days, the Secretary has to send Congress a permanent oversight framework — covering everything from licensing to monitoring AI hardware and cloud usage — and update it at least once a year. The whole program sunsets 7 years after enactment unless Congress renews it.