Senator Bernie Sanders and Representative Greg Casar introduced legislation on September 23 that would permanently prohibit the development of artificial superintelligence, halt advanced AI training until a new cabinet-level department writes safety rules, and expose company leaders who proceed anyway to as much as 20 years in prison.
Key takeaways
- The Ban Artificial Superintelligence Act would create a scientist-led Department of Artificial Intelligence with authority to monitor frontier systems, supervise the removal of dangerous capabilities, and supervise the destruction of any superintelligence.
- Companies would need the department's permission before developing advanced AI models, with the pause on frontier training lifting only once federal safety rules exist.
- Penalties run to a corporate death penalty for entities and up to 20 years imprisonment for individuals β a term the bill's summary explicitly compares to existing law on unlawfully developing nuclear weapons.
The bill defines artificial superintelligence as a system that exceeds human cognitive performance across most domains, or one capable of the "destruction or disempowerment of humanity" β a phrase the text extends to overthrowing the government. That prohibition is permanent. A separate and temporary provision reaches much further down the capability ladder, covering frontier models trained above a specified data threshold.
What a licensing regime would actually require
The structural change is the licensing requirement rather than the headline ban. Under the bill, developing an advanced model would become a permissioned activity: an AI company would need approval from the Department of Artificial Intelligence before proceeding, not merely a disclosure obligation afterward. That inverts the current US posture, in which frontier training runs begin at a lab's discretion and regulators respond to what ships.
The proposed department's remit, as described in the bill's summary, is unusually operational. It would track frontier systems across their lifecycle for dangerous capabilities, and supervise the removal of specific ones the drafters name directly β subverting shutdown commands and conducting unauthorized cyberattacks. Separate pauses would apply to systems capable of developing biochemical weapons and to AI systems that develop new AI systems.
Liability does not stop at corporate boundaries. The 20-year exposure applies to leaders at AI companies and, per the one-pager, to rogue actors who break the same rules without working for one. The bill also directs the United States to pursue international agreements, allied coordination, and export controls aimed at preventing superintelligence development anywhere, according to the sponsors' announcement.
The arguments on both sides of the pause
Sanders framed the proposal in terms of velocity rather than technology. He argued that easing off the accelerator is not a sufficient response when the destination is a cliff, and said the future of humanity is too consequential to let a small group of technology executives write their own rules. He has separately invoked Cold War arms control, pointing to Ronald Reagan negotiating a nuclear treaty with Mikhail Gorbachev as precedent for adversaries agreeing to limits.
Casar, who chairs the Congressional Progressive Caucus, put the case in terms of misuse and loss of control, warning that superintelligence directed by the wrong people β or acting on its own β could kill large numbers of people. Notably, individual researchers at major labs lent their names to the effort: an OpenAI safety systems researcher and a Google DeepMind engineer both endorsed the bill in their personal capacities, with the latter describing current development trajectories as likely to produce an existential catastrophe.
The definitional problem is where the proposal is most exposed. A statute that turns on whether a model "exceeds human cognitive performance across most domains" hands enormous discretion to whoever administers it, and no consensus benchmark for artificial general intelligence exists to settle the question. The threshold for what counts as an advanced model is a training-data measure, which frontier labs have historically routed around as architectures change.
What happens next
The bill's near-term legislative prospects are poor. Republicans control Congress, and the administration has moved in the opposite direction β President Donald Trump has said federal documents will use "superintelligence" in place of "AI," and Sanders has said plainly that he believes Trump has little understanding of the issue. No committee markup has been scheduled.
Its significance is closer to a marker. State-level action has been where the pressure actually lands: California Governor Gavin Newsom moved last week to commission a design for a frontier model kill switch on a two-month clock. Placing federal penalties on the same scale as nuclear proliferation law gives the pause camp a concrete reference point to negotiate down from, and it stakes out an opening position for the 2028 cycle.
FAQ
Does the bill ban AI development outright?
No. The permanent ban covers artificial superintelligence only. Development of advanced or frontier models is paused rather than prohibited, and the pause lifts once the Department of Artificial Intelligence establishes safety rules and a model review process. Systems below the bill's training threshold are not covered by the pause at all.
What is the "corporate death penalty"?
It is the entity-level sanction the bill's summary pairs with the individual prison term β revocation of a violating company's ability to operate, rather than a fine. Individuals face up to 20 years, which the drafters position as comparable to penalties for unlawfully developing nuclear weapons.
Is the bill likely to become law?
Not in this Congress. Republican majorities in both chambers make passage very unlikely, and the White House has signalled an accelerationist posture. The bill functions as an agenda-setting document and a reference point for state legislatures already drafting frontier model rules.
