Banning superintelligence: the bill that moves the debate from voluntary codes to the criminal code

Until now, American proposals on artificial intelligence have regulated uses: who may do what, with which safeguards, under whose oversight. The bill introduced on 23 September 2026 by Senator Bernie Sanders (independent, Vermont) and Representative Greg Casar (Democrat, Texas), who chairs the Congressional Progressive Caucus, does something different: it prohibits. 'Artificial superintelligence' — defined as a system that exceeds human cognitive performance in most domains, or that has capabilities sufficient to destroy or disempower humanity, including by overthrowing the federal government — could be neither developed nor deployed. Alongside the ban, a cabinet-level Department of Artificial Intelligence would be created, responsible for safety and for enforcing the prohibitions. Development of advanced AI would be paused until that body had set rules and review procedures, and according to the Associated Press, deploying advanced systems would then require federal approval. Some capabilities would be stopped immediately, without waiting for the rules: the press releases cite the development of biochemical weapons and the ability of an AI to build new AIs in place of humans.

What really sets it apart is the penalties. For companies, the bill provides for the 'corporate death penalty', meaning dissolution; for individuals, up to twenty years in prison. The sponsors name their benchmark openly: the penalties for illegally developing nuclear weapons. Internationally, the bill sets as a US objective agreements between countries, coordination among allies and tools such as export controls, so that the ban doesn't stop at national borders. "Our bill bans the development of artificial superintelligence and pushes for international agreements so that no one, anywhere, builds AI too powerful for humans to control," Casar said. And Sanders: "When the future of humanity is at stake, we need binding international safety rules, not voluntary standards from the industry."

This is where the problem the press release leaves unsolved begins. The definition of superintelligence is qualitative: there is no compute threshold, no measurable parameter, no date. It is unclear which systems would count as the 'advanced AI' subject to the pause, or how long the pause would last. I should say that I could not read the full legislative text: the Senate website and the summary PDF both returned a 403, so everything above comes from the two official press releases and the AP report. No co-sponsors beyond the two sponsors are on record so far, and according to AP the bill has little chance of passing a Republican-controlled Congress, where there is no agreement even on far narrower rules. The White House's stated position so far opposes greater public oversight. Also speaking at the launch were a researcher from OpenAI's Safety Systems team and a Google DeepMind software engineer, both stressing that they were speaking in a personal capacity. The former summed it up this way: "superintelligence could either go extremely right or extremely wrong."

A law that bans something without being able to say when that thing exists is, in terms of enforcement, a serious problem: an agency would end up setting the threshold case by case, and that threshold is where the legal challenges would pile up. But the interesting effect lies elsewhere, and it doesn't depend on the bill passing. For years the public question has been how much safety companies voluntarily pledge to provide; this bill recasts it as how much they must provide to avoid committing a crime. The difference between a pledge and a criminal charge is not one of degree: in the first case the company decides, in the second the law does.

— Olya