On September 23, 2026, Sen. Bernie Sanders (I-Vt.) and Rep. Greg Casar (D-Texas) introduced the Ban Artificial Superintelligence Act. Earlier this month, the argument over slowing AI was a matter of CEO essays and interviews, with Mark Zuckerberg and Jensen Huang publicly rejecting a coordinated slowdown. This is the first time that argument has taken the form of proposed federal law.
The bill does four things, according to the sponsors' own releases:
- A permanent ban on developing and deploying artificial superintelligence, defined as AI that "exceeds human cognitive performance and capabilities across most domains, or has sufficient capabilities to destroy or disempower humanity, including by overthrowing the federal government."
- A temporary pause on advanced AI development that lasts until a new regulator has written safety rules and a model review process. There is no fixed end date; the pause ends when the agency is operational.
- A cabinet-level Department of Artificial Intelligence to monitor frontier systems across their lifecycle, oversee removal of dangerous capabilities, and oversee destruction of any superintelligent system.
- Penalties the sponsors describe as a "corporate death penalty" for companies, and up to 20 years in prison for individuals, which the sponsors compare to penalties for unlawful nuclear weapons development. It also directs the US to pursue international agreements, allied coordination and export controls.
The sponsors' releases name no additional cosponsors, and the administration has publicly opposed a coordinated slowdown, so near-term passage looks unlikely.
I'd read this bill as a marker rather than a statute, and the useful thing it marks is how hard the definition problem is. The ban's first prong, "exceeds human cognitive performance and capabilities across most domains", has no test attached. No benchmark suite anyone agrees on can say when a model crosses "most domains", and the evals labs publish are exactly the numbers they would be motivated to shade if a 20-year sentence hung on them. The second prong, capability to "disempower humanity", is even less measurable ahead of time. A prohibition with criminal penalties and an unmeasurable trigger either never fires or gets decided by whoever writes the evaluation. That is why the pause mechanism matters more than the ban: it hands the real definitional work to an agency that does not exist yet. If this debate gets anywhere, it will be through that eval-and-review machinery, which is also where anyone who takes AI risk seriously, on either side of the slowdown argument, should be putting effort.
The slowdown debate has moved from CEO statements to proposed law, but the bill's ban hinges on a capability definition no current evaluation can test -- so the fight that matters is over who builds the measurement and review process.
Sources
- Casar, Sanders Introduce Legislation to Create New Federal Agency to Ban Artificial Superintelligence (Rep. Casar)
- Ban Artificial Superintelligence Act one-pager (Sen. Sanders)
- Bernie Sanders and Greg Casar propose AI 'superintelligence' ban with a 20-year jail penalty (NBC News)
- Sanders, Casar call for artificial superintelligence ban (The Hill)