California Democratic Rep. Ro Khanna leaves the U.S. Capitol after voting for the last time this week on Thursday, September 3, 2026.
Tom Williams | Cq-roll Call Inc. | Getty Images
Representative Ro Khanna, a Silicon Valley Democrat, plans to introduce a bill to regulate artificial intelligence that would include strict liability standards and a ban on “recursive” self-improving AI until the government puts safeguards in place.
The bill comes amid calls for the U.S. government to regulate AI after top officials warned the technology could spiral out of control. While Congress has propped up the economy, it has done little to regulate the technology, which can be devastating if done wrong.
“In fact, there is a risk of civilization ending,” Khanna said in an interview with CNBC on Monday. “There is a safety risk of loss of control, and there is also a risk of misuse, both of which need to be taken seriously.”
Khanna’s bill, a summary of which was shared exclusively with CNBC, is called the Human Control over AI Act, and would prohibit models that recursively self-improve or autonomously change their core purpose of containment or shutdown control until federal guardrails exist and agencies approve the activity.
The bill also includes OpenAI, Anthropic, google DeepMind and xAI are gaining attention as safety concerns rapidly increase.
The agency will be tasked with developing safety regulations, a licensing system to approve model training and deployment, auditing of frontier AI models, and setting security standards for continuous testing and effective human control.
It would also require all frontier labs to have independent auditors and report directly to authorities. The agency also plans to create standards for sandbox testing environments. Air gap, or isolate your test environment from the internet. kill switch. It also includes controls that prevent models from escaping the lab setting or roaming on the Internet. It would also be responsible for regulating the use of advanced chips used by Frontier Laboratories, including mechanisms to monitor or control their use.
The bill would also create new legal penalties if AI fails.
It criminalizes the introduction of AI models that result in the destruction of civilians as a “crime against humanity.” The bill would also impose criminal penalties on employees of AI companies who disable safety equipment, kill switches, logging or containment systems, or intentionally deploy unauthorized systems.
AI companies will also be required to carry extensive liability insurance to release models.
The bill also requires the government to pursue enforceable agreements and targeted export controls to prevent China and other adversaries from developing dangerous AI systems.
Khanna said the bill represents “the most comprehensive AI safety legislation” ever proposed. He said this is not a request from executives at major AI companies, but rather is modeled after conversations with AI safety organizations such as the independent nonprofit Model Assessment and Threat Institute, Machine Intelligence Institute, and Palisade Research.
“We want this to be a model for giving voice to the AI safety community,” Khanna said. “I think they were unfairly dismissed as science fiction and not taken seriously enough…They, the AI safety community, need to be leading this discussion, not Frontier Institute executives.”
The bill now joins a number of other AI safety proposals, including the Frontiers Act, a bipartisan proposal from Rep. Jay Obanolte, R-Calif., and Rep. Lori Trahan, R-Mass., that would put independent auditors at Frontier Institute and allow the government to shut down models that pose potentially catastrophic risks.
The House bill is not expected to be voted on until after the midterm elections, and the Senate is expected to leave Washington after this week and not return until after the elections.
