Bernie Sanders wants the government to decide which kinds of artificial intelligence must never be built. The most consequential part of his proposal is buried in the definition.
On September 3, Sanders and Representative Greg Casar announced the Ban Artificial Superintelligence Act. Their plan combines a permanent prohibition with a temporary development pause, overseen by a new federal regulator. The announcement describes forthcoming legislation. The materials reviewed for this article establish a proposal, not an enacted ban. Sanders’s official announcement
That distinction matters. A political demand can reshape the debate long before it changes the law. It deserves scrutiny on its actual terms.
The definition is the story
The sponsors’ one-page summary defines artificial superintelligence to include systems that “match or exceed human cognitive performance” across a broad range of domains or tasks. It also includes systems that could easily be modified to do so. A separate branch covers capabilities sufficient to disempower humanity, including undermining the US government. Official proposal summary
Read that first branch carefully. Matching humans is enough; vastly surpassing them is unnecessary. The requirement spans multiple domains, so this does not straightforwardly describe a calculator or a chess engine. But it could reach considerably further than readers might infer from a ban on superintelligence. The eventual legislative text would need to make that boundary intelligible.
Which humans provide the comparison: an average adult, a trained professional, or the best specialist? How many tasks constitute a broad range? Does a system qualify when it answers questions successfully, or only when it can perform the work reliably? Those are questions the published summary leaves unresolved.
This is a practical measurement problem. NIST’s 2024 generative AI risk profile warns that performance on tests designed for humans does not establish a system’s validity or reliability in the corresponding domain. It also describes gaps between controlled evaluations and real-world use. A licensing-exam score, in other words, cannot carry the entire burden of a legal definition. NIST generative AI risk profile, Appendix A.1.4
Nor does uncertainty establish safety. NIST’s broader risk framework explicitly cautions that difficulty measuring a risk does not make it inherently high or low. That cuts both ways: developers cannot convert incomplete evidence into permission to do anything, and lawmakers cannot convert an evocative label into a dependable test. NIST AI Risk Management Framework, section 1.2.1
Who gets to draw the line?
The proposal’s strongest challenge is institutional. Its planned cabinet-level agency would monitor frontier systems and oversee the removal of dangerous capabilities and the destruction of prohibited systems. The announcement also outlines substantial penalties, including prison sentences of up to 20 years for individuals who violate or circumvent its restrictions. Official announcement
Our view: the public has a legitimate interest in deciding when a company’s experiment creates unacceptable risks for everyone else. That authority needs machinery behind it—qualified investigators, access to evidence, clear intervention powers and a way to challenge decisions. Giving a regulator an imposing name would accomplish little if it remained dependent on the companies it supervised.
The temporary pause raises another design question. The sponsors would keep it in place until the new regulator had established rules and a review process. That makes institutional readiness the release condition. Readers should therefore look for measurable requirements, funding and implementation milestones in the eventual bill. An undefined end condition would be a consequential policy choice in its own right. Official proposal summary
International enforcement is harder still. A research proposal by Aaron Scher and colleagues offers a useful illustration: it couples limits on training with chip tracking and verification, centered on US–China cooperation. Its authors acknowledge both the absence of sufficient political will and the possibility that technical progress could weaken their approach. This is a proposed framework, not an existing agreement or proof that a global prohibition would work. Scher and colleagues, revised May 2026
A proposal still owes us the details
Readers should also keep this initiative separate from the data-center moratorium Sanders announced with Alexandria Ocasio-Cortez on March 25. That earlier proposal targets infrastructure expansion pending safeguards covering safety, workers and communities. The September proposal addresses the development and deployment of specified AI systems. The political concerns overlap; the mechanisms differ. March announcement
What deserves attention next is the legislative text: precise capability thresholds, who performs evaluations, how decisions are reviewed, and what existing systems would face. Until those details are available, sweeping claims about an imminent shutdown run ahead of the evidence.
Sanders has put an uncomfortable question squarely into politics: who gets to decide that a more capable machine is worth the risk? His proposal now needs to answer an equally demanding question—how that decision could be enforced without turning an uncertain scientific boundary into an arbitrary legal one.
How this article was made
This analysis was prepared using AI-assisted source research and drafting, with checks against the linked public documents. Interpretive judgments are distinguished from the sponsors’ claims. Publication was authorized by the publisher. No original interviews or separate human fact-check are claimed.
