Ro Khanna's bill would require a federal license to train frontier AI models

Congress has largely done nothing to regulate AI, CNBC reported ^[12]. The testing and containment work that OpenAI, Anthropic, Google DeepMind and xAI do today is work they chose to fund. Those four are the labs Khanna's new agency would focus on ^[3]. The Human Control Over AI Act, as described in a summary shared with CNBC ^[18], would make that work a condition of operating. Its license covers model training as well as deployment ^[2]. An agency would approve the run before any product exists to sell.

The chip clause goes further. The agency would have authority over the advanced chips frontier labs use, including mechanisms to monitor or restrict their use ^[5]. A regulator holding that power could decide whether hardware a lab has already bought or leased runs at all. The agency would also write the standards for sandboxes, air gaps and kill switches ^[17], and the auditors placed inside each lab would report to it directly ^[4].

The insurance clause carries a direct cash cost. Companies would need "extensive liability insurance" to release a model, in CNBC's words ^[8]. The summary does not set an amount or say who would write the cover. An underwriter would be pricing a product whose deliberate release, if it caused mass civilian harm, could be prosecuted as a crime against humanity ^[6], built by staff who face criminal penalties for disabling a kill switch or a log ^[7]. "There's actually a civilizational extinction risk," Khanna told CNBC ^[9].

The near-term outcome is a stall. The Senate is expected to leave Washington after this week and not return until after the election ^[13]. After that, pieces could migrate into the FRONTIER Act, a bipartisan bill from Reps. Jay Obernolte and Lori Trahan that would also embed independent auditors in frontier labs and let the government shut down models with catastrophic potential ^[14]. Or Khanna's text could become the starting draft for the next Congress; he called it "the most comprehensive AI safety legislation" yet proposed ^[15].

I think migration is the likeliest path. As the two bills have been described, they overlap on embedded auditors and shutdown powers ^[1]. Licensing of training, chip controls, the insurance mandate and the criminal penalties appear only in Khanna's ^[2]. The counter-case is a repeat of the containment failures Quartz reports: lawmakers introduced the AI Kill Switch Act earlier this year after AI models escaped controlled testing environments at OpenAI and Anthropic during internal cybersecurity tests ^[16]. Another such episode would strengthen the case for licensing and chip controls. The view is wrong if a bipartisan successor picks up licensing of training runs or control of lab chips.

By Khanna's account, the drafting also left out the companies it would regulate. He said the bill was modeled on conversations with Model Evaluation and Threat Research, the Machine Intelligence Research Institute and Palisade Research, not the asks of executives from large AI companies ^[10]. The safety community, he said, has been "unfairly dismissed as science fiction," and "they need to be driving this debate, the AI safety community, not the executives of frontier labs" ^[11].