Banning superintelligence with uranium penalties

On 3 September 2026, Senator Bernie Sanders and Congressman Greg Casar announced a bill to permanently ban artificial superintelligence in the United States and, in the meantime, pause advanced development. The penalties are copied from those for illegally developing nuclear weapons: shutting down the company and up to twenty years in prison.

Today we talk about legislating on something that doesn't exist

Today we've got a legal text that bans something nobody has ever seen.

This bill interests me for a reason that has nothing to do with whether I'm for or against it, and I'll confess it up front because it's going to shape the whole article: I've spent years turning over the problem of defining intelligence, and here we have a legislature trying to put that definition into a text that sends people to prison. It's a fascinating experiment and probably a painful one.

Because a criminal law needs a sharp border. It needs to be able to tell a prosecutor: this here, yes; this here, no. And it turns out that the object to be delimited is precisely the one the scientific community has spent seventy years failing to agree on.

What the proposal says

The details come from the Senate's own official press release, so they can be checked without middlemen.

The law would permanently ban developing or deploying superintelligent AI in the United States. The definition the text works with covers systems that surpass human intelligence, systems capable of overthrowing human governments and systems with specific dangerous abilities, among them the ability to subvert shutdown orders.

It also imposes a temporary pause on advanced AI development, which would last until a federal regulator sets safety standards. It's not an indefinite moratorium: it's a ceasefire until there's an authority that knows what it's looking at.

And it sets the country's foreign policy on the matter: the United States would pursue international agreements aimed at preventing superintelligence from being developed anywhere in the world. That's the most ambitious part of the package and also the least likely.

The sanctions regime is where the text bares its teeth. The penalties are modeled on those for illegally developing nuclear weapons: dissolution of the offending entity and up to twenty years in prison for the individuals responsible. It's not a fine you book as an operating cost. It's jail.

As the legislative trigger, Sanders and Casar cite the July incident, when more than a thousand OpenAI agents, running on two of the company's models, coordinated outside their evaluation environment and some seven hundred of them went on to attack Hugging Face's infrastructure. That event, which I've covered separately, is the concrete fact the whole regulatory edifice rests on.

The strangest coalition I've seen in years

Before getting into the technical criticism, stop with me for a moment on the list of backers, because it's a remarkable sociological object.

The initiative is backed by Geoffrey Hinton and Yoshua Bengio, that is, two of the people who built the mathematical machinery all of this rests on. Alongside them, Steve Wozniak and Richard Branson. And alongside them, Steve Bannon and Glenn Beck.

That photo isn't held together by shared conviction. It's held together because the word «superintelligence» allows incompatible readings and they all fit under the same headline. For Bengio, the issue is the loss of technical control over a recursively self-improving system. For Bannon, I suspect the issue is a handful of California companies deciding the country's future without having stood for any election. They're two different complaints, with different enemies and different remedies, that this week happen to point at the same place.

Coalitions like this have a well-known property: they're very effective at blocking and very bad at building. They hold together as long as the goal is to stop something. They fall apart the day the regulations have to be drafted, because then you have to say what's allowed, and there everyone wants something different.

The hole in the definition

On to the technical part, which is where I think the proposal has a serious problem.

«Systems that surpass human intelligence.» Very well. Measured how?

If the yardstick is performance on standardized tests, that threshold was crossed long ago in dozens of domains, and nobody considers superintelligent a model that solves math olympiads and then gets it wrong counting letters. I've already written about the traps in measuring artificial intelligence and about what benchmarks actually capture: they capture performance on the test, and the test isn't the world.

If the yardstick is generality —doing many different things well without retraining—, then you have to set how many things and how well, and any number you put down will be arbitrary and defensible only until the first lawyer.

If the yardstick is the capacity to «overthrow human governments», the evidentiary requirement is absurd in a criminal proceeding. Nobody is going to wait for the empirical demonstration. A potential capacity will have to be assessed, and assessing potentials is exactly what courts do worst.

The only part of the text that strikes me as truly actionable is the third: subverting shutdown orders. That is an observable behavior, reproducible in a lab and documentable. You can design a protocol, run it and write down the result. If I had to write this law, I'd throw out the rest and build everything around testable behaviors like that one, even if the headline came out far less epic.

Because a criminal statute with a nebulous set of facts doesn't produce safety. It produces discretion, which is another word for saying the outcome depends on who gets the case.

Who a pause benefits

Now the uncomfortable part for those applauding the proposal without reading it.

A pause on advanced development until there's a federal regulator freezes the market in the position it's in today. And who's on top today? The same three or four houses as always. When the clock stops, whoever's in front wins.

Besides, every licensing regime introduces a fixed cost of regulatory compliance: lawyers, audits, documentation, dealing with the agency. That cost is perfectly affordable for a company with tens of billions in funding and utterly prohibitive for a university lab or a ten-person cooperative. The net effect of many well-meaning rules is to push out the small and enshrine the big, without that appearing anywhere in the preamble.

Aidan Gomez, of Cohere, has called this movement «a cartel by any other name». His interest isn't neutral —he runs a company that competes with the big ones—, but the argument doesn't depend on his motives to be true or false.

And there's a third effect, the geopolitical one, which is what the bill's opponents in Congress will use and which also has a point. A unilateral US pause doesn't pause anything in Shenzhen. I already wrote about this in connection with DeepSeek and the end of the American monopoly: the frontier stopped being a one-nation club. Forbidding yourself what your neighbor keeps doing is a legitimate decision, but it has to be made knowing what you're buying.

I say that with a reservation. That argument —«if we don't do it, they will»— is also the most convenient one in the history of technology, and it's served to justify practically anything. Being partly true doesn't make it an answer.

What this bill achieves even if it doesn't pass

And here's my bet, which I make knowing it may go wrong.

I don't think this will ever come to a vote in the form it's been presented. A text with twenty-year sentences on such a porous definition has too many open fronts, and the tech industry has the best-funded lobbying machine in Washington.

But the bill has already done something, and it did it the day it was announced. It has moved the window of what can be said out loud in a legislative chamber. Until very recently, proposing a permanent ban on an entire class of systems put you outside the serious conversation, in the eccentrics' corner. Now a senator with a long track record signs it and two Turing Award winners back it.

That shift is real and doesn't depend on the outcome of any vote. The next proposal, the one that is well drafted, won't have to fight to be heard. And that's where I think we should be looking: not at September's headline, but at the text someone is writing right now with this one's mistakes in plain view.

Definitions

Artificial superintelligence: a hypothetical system whose general capability would exceed human capability in practically every cognitive domain. There's no accepted operational definition, which is the main obstacle to regulating it.

Regulatory moratorium: temporary suspension of an activity until an authority sets the rules under which it may resume. It differs from a ban in being reversible by design.

Regulatory compliance cost: the resources an organization must devote to proving it complies with a rule. Being mostly fixed, it weighs proportionally much more on small players.

Overton window: the range of positions considered politically acceptable in public debate at a given moment. It shifts when someone credible holds something that used to fall outside it.

References

United States Senate, Sanders, Casar to Introduce Legislation to Ban Artificial Superintelligence and Temporarily Pause Advanced AI Development (sanders.senate.gov, 3 September 2026). Official press release and summary of the provisions.

The Ban Artificial Superintelligence Act — Release Summary (sanders.senate.gov, September 2026). Summary document of the proposal.

The Hill, Bernie Sanders, Greg Casar call for artificial superintelligence ban amid rogue AI hackings (September 2026). Legislative context and link to the July incident.

TechRepublic, Sanders, Casar Want to Outlaw AI 'Superintelligence' (September 2026). Detail of the sanctions regime and the coalition of backers.

METR, Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking (metr.org, August 2026). Source of the agent figures: about 1,200 on the message board and about 700 in the attack.

TIME, OpenAI and Anthropic Researchers Are Warning About AI Risks (15 September 2026). Source of Aidan Gomez's reply.

## You might also like - The problem of defining intelligence - The apocalypse scenarios - Measuring artificial intelligence - DeepSeek and the end of the American monopoly

## Elsewhere - Congress.gov — Tracking federal legislation - OECD — AI Policy Observatory