On 3 September, Senator Bernie Sanders (I-Vt.) and Representative Greg Casar (D-Texas) introduced the Ban Artificial Superintelligence Act, a bill that would do exactly what the name says: permanently outlaw the development and deployment of superintelligent AI in the United States, and pause frontier development in the meantime until a federal regulator writes safety rules. The penalties are not symbolic. Individuals could face up to 20 years in prison. Companies could face what Sanders’s office calls the “corporate death penalty”, a forced shutdown. The sponsors compare the whole framework to the laws governing the unlawful development of nuclear weapons.
That comparison tells you how the bill’s authors want you to think about this. Not as a tech-policy squabble over training data, but as a category of thing so dangerous that the state reserves the right to end your company for pursuing it.
What the bill actually bans
The Act draws a line at “superintelligence”, broadly, AI that outperforms humans across essentially all cognitive tasks. Below that line, ordinary AI tools carry on. Above it, development stops. The temporary pause on advanced (frontier) development would hold until a federal regulator certifies that safety and controllability standards exist and are being met. In plain terms: the government wants a hand on the brake before anyone builds a machine cleverer than the people trying to control it.
Sanders framed the urgency in his announcement: “Nearly every day, there is a frightening new story about how Big Tech companies are losing control of the technology they are developing, with potentially cataclysmic results.” Casar put a clock on it: “In just four years, we have gone from the first version of ChatGPT to AI models so powerful they cannot be properly controlled.”
Who is backing it
The striking thing about the coalition is that it does not map onto the usual political lines. The broader campaign for a superintelligence ban has drawn signatures from people who agree on almost nothing else: AI pioneers Geoffrey Hinton and Yoshua Bengio (both of whom helped invent the field), Apple co-founder Steve Wozniak, Virgin’s Richard Branson, and, from the other end of the spectrum entirely, Steve Bannon and Glenn Beck. When Hinton and Bannon are on the same petition, something unusual is happening.
The public, at least in polling, is sympathetic. Surveys released alongside the campaign found that 64% of Americans think superintelligence “shouldn’t be developed until it’s provably safe and controllable”, and only 5% want it built as fast as possible. Separately, 82% of voters said they do not trust tech executives to regulate themselves. Whatever you make of the bill, the mood it is riding is real.
The case against
Deferring to the sceptics here matters, because the opposition is not just the industry protecting its book. The loudest practical objection is geopolitical: a unilateral American ban does nothing to stop Beijing or Moscow, and critics warn it would hand them a decisive lead in the one race that arguably matters most. A ban you cannot enforce beyond your own borders is a handicap you impose only on yourself.
The more interesting objection comes from within the safety camp. Gary Marcus, one of AI’s most persistent critics and no friend of the labs, opposes the bill as written. His view is that the instinct is defensible but the instrument is too blunt: “we may need a temporary pause, maybe even one that lasts for a number of years, until such time as we have much better ideas about regulation and alignment.” A pause, in other words, is arguable. A permanent criminal ban on a term nobody can precisely define is a different proposition. Science magazine ran a whole piece on exactly that problem, that experts cannot agree what “superintelligence” even means, which is an awkward foundation for a 20-year prison sentence.
Will it pass? Almost certainly not this year
Here is the sober part. With midterm elections in November 2026, Congress has little appetite to move sweeping technology legislation, and a bill this radical faces long odds regardless of the calendar. The tech industry will spend heavily against it. Realistically, the Ban Artificial Superintelligence Act is a marker, a way of dragging the Overton window and forcing the “should there be a hard limit at all?” question onto the record, rather than a law that lands on the President’s desk any time soon.
That does not make it irrelevant. Bills like this shift the terms of the debate even when they fail. Five years ago the idea of criminalising a category of software carried a 20-year sentence would have sounded absurd. It is now a numbered bill with signatures from Turing Award winners. Whether the Act itself lives or dies, that shift is the story.
For the labs, the immediate effect is reputational rather than legal. It is one more data point in a year where the pressure on frontier AI has moved from op-eds to courtrooms and now to statute books. For the rest of us, it is worth watching less for whether it passes and more for what the vote count, if it ever comes, reveals about how nervous Washington actually is.
None of this is legal or investment advice; it is a summary of a fast-moving policy fight, and reasonable people disagree about all of it.
Related reading on Top Tool Stack: OpenAI Launched GPT-6 Astra and Declared the AGI Era. Steady On. and Anthropic built an AI that does AI safety research.