Policy#AI safety#Model training
Rep. Khanna prepares the Human Control Over AI Act
CNBC reports Rep. Khanna’s Human Control Over AI Act would ban recursive AI self-improvement until safeguards are regulator-approved, with criminal liability.
Per CNBC (relayed by IT之家), Democratic Rep. Ro Khanna — whose district covers Silicon Valley — is preparing to introduce the Human Control Over AI Act: until federal safeguard rules exist and a regulator approves, AI models may not recursively self-improve, nor autonomously modify their own core objectives, containment mechanisms or shutdown controls. Khanna calls it the most comprehensive AI safety legislation yet, saying “there is a risk of civilizational extinction — both from losing control and from misuse, and both must be taken seriously.”
The facts
- The core ban: no recursive self-improvement and no autonomous modification of core objectives, containment mechanisms or shutdown controls before safeguards exist and are approved.
- A new agency: a federal AI safety body would write the rules, license model training and deployment, and audit frontier models; every frontier lab would need independent auditors reporting directly to the agency.
- Technical standards: sandbox test environments, physical internet isolation, emergency kill switches, controls preventing models from running on the open internet outside the lab, and monitoring of advanced chips.
- Liability: AI companies would carry broad liability insurance to ship models; deploying an AI that destroys a civilian population would be prosecuted as a crime against humanity, and employees who disable safeguards, kill switches or logging would face criminal penalties too.
- Sourcing note: the bill is at the “preparing to introduce” stage, not formally filed; the reporting chain is CNBC → IT之家, and the original was not independently verified.
Our take
This is the first serious attempt to put recursive risk into federal statutory text, with clauses down to auditor reporting lines and chip controls. If it becomes the reference debate, the industry prepares more than a compliance department — it answers “who approves the guardrails,” now contested on three fronts: Florida’s injunction motion, LASST’s civil suit, and this bill. The cold read: a Silicon Valley district representative’s bill is heavy on symbolism and far from law — but OpenAI’s own training pause and the scrapped Astra 6.1 keep handing it evidence.