Silicon Valley Democratic Rep. Ro Khanna shared a summary of his Human Control Over AI Act with CNBC on 28 September 2026. In an interview he said there is “a civilizational extinction risk,” a loss-of-control safety risk and a misuse risk, and that both must be taken seriously.
The core of the bill, per CNBC, The Next Web and Quartz, is a conditional ban: forbid models that recursively self-improve or that autonomously modify their own core objectives, containment or shutdown controls until federal guardrails exist and a designated agency approves those activities.
The law would create a new federal agency focused on models from OpenAI, Anthropic, Google DeepMind and xAI. Its tasks: write safety rules, a licensing system for training and deployment, frontier-model audits and standards for continuous testing and effective human control.
Independent auditors would be embedded at every frontier lab and report directly to the agency. It would also set standards for sandbox test environments, air gaps, kill switches and controls to keep models from “escaping” lab settings. The agency would regulate advanced chips those labs use, including monitoring or restriction mechanisms.
The criminal chapter is stark: it would treat deployment of AI models that destroy civilian populations as “crimes against humanity”; impose criminal penalties on employees who disable safeguards, kill switches, logging or containment, or who knowingly deploy an unauthorised system; and require extensive liability insurance before models ship.
Khanna called the bill “the most comprehensive AI safety legislation” proposed and said it was modelled on conversations with safety organisations—METR, MIRI, Palisade Research—rather than frontier-lab executives’ asks. He wants to “give voice to the AI safety community,” which he says has been “unfairly dismissed as science fiction.”
The bill joins the bipartisan FRONTIER Act from Reps. Jay Obernolte (R-Calif.) and Lori Trahan (D-Mass.), which would also put auditors in labs and let the government shut down models with catastrophic potential. CNBC says no House bills are expected to get a vote until after the midterms; the Senate may leave Washington until after the election.
The contrast with the 29 September White House pact is sharp: Khanna wants licences, crimes and chip oversight; Trump offers company-chosen audits with no penalties. Both speak of “human control,” but only one creates an agency that can deny training.
For labs, compliance costs would be structural: auditable kill-switch design, log retention, large insurance policies and resident auditors. For startups outside the frontier club, the risk is that the “frontier model” definition widens and captures smaller systems.
Language on China—enforceable agreements and export controls—aligns the bill with the race narrative, but subordinates it to safety: the goal is not only to win, but not to deploy RSI without federal permission.
Observers doubt political viability in 2026. Even so, the text stakes out a pole on the legislative spectrum: alongside Sanders/Casar’s Ban Artificial Superintelligence Act (pause and ASI ban with up to 20-year penalties, per NBC), it shows how far one Democratic faction is willing to go.
Neural Pulse concludes that the Human Control Over AI Act, though unlikely before November, already shifts the debate’s centre of gravity—from voluntary codes to whether recursive self-improvement should be legal without a licence. Labs and boards should already map which pipelines touch autonomous modification of objectives or containment.
