Representative Ro Khanna intends to file the Human Control Over AI Act, legislation that would stand up a federal agency to license and audit frontier models and would forbid recursively self-improving systems until safeguards are in place and that agency signs off on the work.
Khanna, a Silicon Valley Democrat, told CNBC that civilizational extinction risk belongs in the same category as misuse and loss-of-control scenarios. Models that autonomously rewrite their own objectives, containment or shutdown controls would be barred. The agency’s remit would stretch to silicon as well: the high-end chips frontier labs depend on, with mechanisms to monitor how those parts get used.
The body would also write safety rules, run a licensing system for training and deployment, set standards for sandbox testing, air gaps, kill switches and containment, and plant independent auditors inside every frontier lab who report straight to regulators.
The penalties are steep. Deployments that end in the destruction of civilian populations would be charged as crimes against humanity, and employees who disable safeguards, logging or containment systems would face charges of their own. Before any release, companies would have to carry substantial liability insurance. It also directs the administration to press for enforceable pacts and tighter export controls focused on China.
Khanna says the substance came from talks with safety researchers outside the big labs rather than from what those labs asked for. Three groups shaped the text: Model Evaluation and Threat Research, Machine Intelligence Research Institute and Palisade Research. No House bills are expected to get a vote before the midterms.
