US Lawmaker Proposes Bill Banning Recursive AI Self-Improvement Without Government Approval

U.S. Representative Ro Khanna has introduced the 'Human Control Over AI Act,' which seeks to ban AI models from engaging in recursive self-improvement without prior government safety approvals. The proposed legislation would establish a new federal agency to regulate frontier AI models, conduct safety audits, and enforce mandatory licensing for training and deployment. This proposal represents one of the most stringent AI safety regulatory efforts to date in the U.S., placing the guidance of AI policy into the hands of independent researchers rather than tech executives. If enacted, it would introduce strict legal liability—including criminal penalties and required liability insurance—for developers whose models escape human control. The bill requires frontier AI labs to utilize independent auditors reporting directly to the regulatory agency, implement physical air-gapping, install emergency kill switches, and monitor advanced hardware usage. Disabling safety features or knowingly deploying unauthorized high-risk systems would carry severe criminal penalties, including potential charges of crimes against humanity in cases of mass civilian harm.