Artificial Intelligence Risk Evaluation Act of 2025

Key Points

• The bill creates a mandatory government testing program for the most powerful AI systems. Companies that build these systems must hand over their code, training data, and model weights so the Department of Energy can evaluate them for dangerous capabilities before they're released to the public.
• The program specifically looks for AI systems that could be weaponized by foreign enemies, lose human control, undermine civil liberties, or show 'scheming behavior' — where an AI tries to deceive its creators by hiding its true capabilities or attempting to bypass safety controls.
• Companies that refuse to participate or release a powerful AI without passing these tests face fines of at least $1 million per day. No one may deploy an advanced AI system for use in interstate or foreign commerce without complying.
• Within one year, the Department of Energy must give Congress a detailed plan for permanent AI oversight. This could include creating a new agency, requiring licenses, or even nationalizing AI projects if they appear to approach 'artificial superintelligence' that threatens humanity.
• The program is set to last 7 years unless Congress renews it, and applies to any AI system trained using more than 10^26 computing operations — a threshold that currently captures only the largest frontier models from companies like OpenAI, Google, and Anthropic.