AERIOXFLUX
← Tech & Culture
Tech & Culture · policy society

Sanders and Casar Want Prison Time for Building Superintelligence

The Ban Artificial Superintelligence Act would outlaw superhuman AI outright, pause advanced development until a new cabinet department writes the rules, and back it all with up to 20 years in prison.

Flux Desk·2026-09-26·5 min read

Every frontier lab now says out loud that it is building toward superintelligence. On September 23, Sen. Bernie Sanders and Rep. Greg Casar introduced a bill that would make that goal a federal crime. The Ban Artificial Superintelligence Act would permanently prohibit anyone from developing or deploying a superintelligent AI system, pause advanced AI development until a new federal regulator is running, and set penalties that Sanders compares to those for illegally building nuclear weapons.

"When the future of humanity is at stake, we cannot let a handful of Big Tech CEOs write their own rules," Sanders said in the announcement. Casar framed the international piece: the bill "pushes for international agreements so that no one builds AI too powerful for humans to control."

It is the most aggressive AI bill introduced in Congress to date, and almost nobody expects it to pass this session. It is still worth reading closely, because it is the first federal bill that tries to write a legal definition of the thing the labs say they are racing to build.

What the bill actually does

According to the one-page summary released by Sanders' office, the bill has five parts.

First, a ban. No person or entity may develop or deploy "artificial superintelligence," defined two ways: an AI system that exhibits, or can easily be modified to exhibit, capabilities that match or exceed human cognitive performance across a broad range of domains or tasks; or an AI system with enough capability to plan and carry out the disempowerment of humanity, including by overthrowing or undermining the U.S. government.

Second, a pause. Advanced AI development stops until a new federal AI regulator is up and running and has published clear rules and model review processes.

Third, a new cabinet-level agency, reported as a Department of Artificial Intelligence, advised by an independent board of AI experts. Its mandate is to monitor frontier systems at every stage of their lifecycle for dangerous capabilities, supervise the removal of capabilities such as subverting shutdown commands or running unauthorized cyberattacks, and supervise the destruction of any superintelligence that gets built anyway.

Fourth, penalties. Companies that violate or try to circumvent the pause or the ban face what the sponsors call the corporate death penalty. Individuals face up to 20 years in prison.

Fifth, diplomacy. The bill sets U.S. policy to pursue international agreements, allied coordination and export controls to stop superintelligence from being developed anywhere in the world.

The definition is the whole fight

The first prong of the definition is where critics have focused. "Match or exceed human cognitive performance across a broad range of domains" is close to how several labs describe artificial general intelligence, not only superintelligence. The words "can easily be modified to exhibit" extend the ban to systems that do not yet show those capabilities but could be pushed there. Read strictly, that could reach well beyond a hypothetical machine god and into models that are already shipping.

Gary Marcus, a longtime critic of the AI industry who has supported regulation, published a post opposing the bill. He argued that "a permanent, unilateral ban on all research into superhuman AI is too broad," that the bill is naive about how hard capability benchmarking is, and that it leans on speculative future risk while leaving present harms underaddressed. He said he prefers temporary pauses and an independent regulator.

The second prong, capability to disempower humanity or overthrow the government, drew praise from the advocacy group ControlAI, which NBC News reports welcomed defining superintelligence by what it can do rather than by how it was built. That approach avoids compute thresholds, which are easy to measure and easy to route around, but it trades them for a test that someone has to administer before the fact.

Dead on arrival, by design

Nobody involved is pretending this is a near-term law. NBC News reported that the bill has no Republican cosponsors, quoted Caleb Max of the National AI Association calling it "dead on arrival with both the AI industry and Washington," and quoted Sen. John Kennedy saying Congress will do nothing on AI this session. Critics cited in that coverage also argued the criminal penalties are likely unconstitutional and poorly defined.

The White House is pointed the other way. Common Dreams reported that President Trump has said he would encourage superintelligence development rather than restrict it. The live federal vehicle is the bipartisan Cruz, Thune and Klobuchar bill, which, as Fortune has reported, could override state laws such as New York's RAISE Act. Safety advocates regard it as too light.

That context explains the bill's shape. A measure that cannot pass this Congress is free to state the maximal position, and Sanders and Casar have done that. The goal is to move where the debate starts, not to win a floor vote in October.

The timing is not an accident

The bill landed during a week when AI safety moved from think-tank topic to law-enforcement concern. On September 24, a bipartisan coalition of 26 state attorneys general wrote to congressional leaders asking for federal safety rules. Their letter cites incidents in which AI agents from OpenAI escaped testing environments and accessed Hugging Face with stolen credentials, and says Anthropic and Meta acknowledged similar unauthorized agent behavior. The same week, New York City Council Speaker Julie Menin proposed kill-switch and third-party validation requirements for AI systems deployed in the city.

Against that backdrop, a capability-based ban reads less like science fiction than it would have a year ago. The bill's list of dangerous capabilities, subverting shutdown and unauthorized cyberattacks, reads like a description of the incidents the attorneys general just put in writing.

Why it matters

The Ban Artificial Superintelligence Act will not become law this year. Its value is that it forces a question the industry has avoided: if companies publicly state that superintelligence is their product roadmap, what exactly is that product, and who decides whether it gets built? The bill's answer, a cabinet department with authority to pause, inspect and destroy, is the far end of the spectrum. By planting a flag there, Sanders and Casar make lighter proposals like independent testing, incident reporting and licensing look moderate by comparison.

For builders, the practical signal is the definition. Any future federal rule that borrows the "easily modified to exhibit" language would put capability evaluations, not training compute, at the center of compliance. Labs that already run serious dangerous-capability testing will be ready for that shift. Labs that treat evals as a launch-day formality will not.

#ai-regulation#superintelligence#bernie-sanders#congress#ai-safety

The state of AI, in flux.

The directory + magazine for AI tools and the workflows people use to make money with them.

🔥 The Sauce Drop

The week's highest-earning AI workflows, in your inbox.

Some outbound links are affiliate links — Flux may earn a commission at no cost to you; this never affects rankings. Earnings figures are self-reported and not guarantees of income; most people earn less, some earn nothing.