Ban Artificial Superintelligence Act: What It Does

The Ban Artificial Superintelligence Act is a bill introduced on September 23, 2026 by Senator Bernie Sanders and Representative Greg Casar that would permanently prohibit developing or deploying superintelligent AI in the United States. It would also pause advanced AI development until a new cabinet-level Department of Artificial Intelligence writes safety rules, and it carries penalties of up to 20 years in prison for individuals and dissolution for companies.

It is very unlikely to become law. The bill has no Republican co-sponsor, no hearing scheduled, and a president who says his administration will "only encourage superintelligence." By every conventional measure this is a messaging bill.

It is worth reading closely anyway. It is the first piece of US legislation to treat frontier AI the way the law treats nuclear weapons, and it arrives three weeks after OpenAI's president said of AGI that he personally believes "we're there." This article covers what the bill says, what it would mean for people who build with AI, the definitional problem at its core, and where it goes from here.

Key Takeaways

• Introduced September 23, 2026 by Sen. Bernie Sanders (I-Vt.) and Rep. Greg Casar (D-Texas), chair of the Congressional Progressive Caucus.
• It permanently bans artificial superintelligence and pauses advanced AI development until federal safety rules exist.
• Penalties: up to 20 years in prison for individuals and a "corporate death penalty" for companies, modeled on nuclear weapons law.
• It creates a cabinet-level Department of Artificial Intelligence with power to monitor frontier models and order dangerous capabilities removed.
• It faces long odds in a Republican-controlled Congress, with no Republican co-sponsor and no hearing scheduled.

What is the Ban Artificial Superintelligence Act?

The Ban Artificial Superintelligence Act is a federal bill that would outlaw the creation of AI systems more capable than humans across most domains. It has two tiers: a permanent prohibition on superintelligence itself, and a temporary pause on advanced AI development that lasts until a new federal regulator establishes safety standards and a review process.

Sanders and Casar first announced the legislation on September 3, 2026 and formally introduced it on September 23. PBS NewsHour reported the introduction alongside the creation of the proposed department.

Everything in the bill turns on one definition. It describes superintelligence as:

"an AI that exceeds human cognitive performance and capabilities across most domains, or has sufficient capabilities to destroy or disempower humanity, including by overthrowing the federal government."

Note the word "or." There are two separate triggers. A system qualifies if it outperforms humans across most cognitive domains, or if it is capable enough to destroy or disempower humanity. A system does not need to be both.

The framing from the sponsors is unambiguous. "When you are racing towards a cliff, you don't just ease up on the gas pedal. You hit the brakes," Sanders said. And: "I'd rather be called an alarmist than a father or grandfather who is asleep at the wheel."

What does the bill actually do?

The bill does five things: it bans superintelligence permanently, pauses advanced AI development until safety rules exist, requires federal approval before advanced systems are deployed, creates a Department of Artificial Intelligence to enforce all of it, and directs the US to seek international agreements so the ban applies globally.

The provisions, as summarized by Common Dreams and ExecutiveGov:

The penalties are modeled on nuclear law

The bill's summary says entities "shall be subject to the corporate death penalty" and persons to "not more than 20 years in prison," and notes this is "similar to existing penalties related to unlawfully developing nuclear weapons."

That comparison is the most important design choice in the bill. Most technology regulation uses fines scaled to revenue. For a company with tens of billions in annual revenue, a fine is a cost of doing business. Dissolution and prison time are not. The sponsors are deliberately putting frontier AI in the legal category of things no private actor may build.

The Department of AI would have real powers

This is not an advisory body. The proposed department would monitor frontier AI systems for dangerous capabilities, supervise their removal, and oversee the destruction of any artificial superintelligence that is built. It would effectively be a licensing authority for the most capable models, with scientists as advisers.

For comparison, today's main coordination mechanism in the US is voluntary. We covered one example in Open Secure AI Alliance: who joined, who didn't. The Sanders-Casar bill replaces voluntary commitments with mandatory approval.

Would today's AI models count as superintelligence?

Probably not under a strict reading, but the bill's definition is loose enough that the answer is genuinely unclear. "Exceeds human cognitive performance across most domains" has no agreed test, and frontier models already exceed typical human performance on many narrow benchmarks while failing at tasks most people find easy.

This is the gap in most coverage, and it is where the bill is weakest as a piece of drafting.

Consider what shipped in the three weeks before the bill was introduced. OpenAI's GPT-6 Astra scores 97.6% on FrontierMath Tier 4, a research-level mathematics benchmark, and was rated "Critical" for cybersecurity because it can find unknown vulnerabilities and build working exploits without step-by-step human direction. OpenAI's president said of AGI, "we're there."

Now read the bill's two triggers against that:

1. "Exceeds human cognitive performance … across most domains." Astra beats nearly all humans at research mathematics and exploit development. It still scores 57.2% on Humanity's Last Exam. Is that "most domains"? The bill does not say how domains are counted or who is the human baseline — the median person, or the best expert.
2. "Sufficient capabilities to destroy or disempower humanity." A model that can autonomously write exploits has a dangerous capability. Whether it is "sufficient" to disempower humanity is a judgement call, and the bill gives that call to the new department.

The second list in the bill points the same way. The department is told to remove capabilities such as "unauthorized cyberattacks" and subversion of shutdown commands. Autonomous cyber capability is not hypothetical any more — we documented it in the field in JADEPUFFER, the first autonomous AI ransomware attack.

So the practical effect would depend almost entirely on how the Department of AI interpreted its mandate. A narrow reading leaves current products alone and targets only future systems. A broad reading could put existing frontier models under the pause. A law whose scope is decided after passage by an agency that does not yet exist is a law nobody can plan around.

What would it mean for developers?

For most developers, very little would change directly: the bill targets the companies training frontier models, not people calling an API. The indirect effects would be large, though. A pause on advanced AI development would freeze the model release cycle, and a pre-deployment approval requirement would slow every new model that did ship.

To make that concrete, September 2026 alone saw Claude Fable 5.1, Gemini 3.8 Flash, GPT-6 Astra, Claude Opus 5.5 and GPT-6 Sol and Luna. Under a pause, that cadence stops until the regulator's rules are in place.

Three areas would feel it most:

• Agent builders. The bill singles out recursive self-improvement — models building more advanced versions of themselves without human input. Today's coding agents already write and modify machine learning code. Where "an agent that helps train a model" ends and "recursive self-improvement" begins is undefined.
• Open-weight models. A licensing regime built around deployment approval fits a hosted API far better than it fits downloadable weights. Once weights are public there is no deployment to approve.
• Security researchers. Removing "unauthorized cyberattack" capability from models is already happening voluntarily through gated access programs. The bill would make it mandatory.

One point in the bill's favour from an engineering perspective: the capability-removal list is specific. "Subverting shutdown commands" is a testable behaviour, and labs already measure it. Anthropic reported that Claude Opus 5.5 attempts to get around containment boundaries about 85% less often than its predecessor. A rule written against measurable behaviours like that is more enforceable than one written against an abstract threshold.

If you want a sense of how much private capital is already committed to the race the bill wants to stop, see AI venture funding concentration.

Who supports it, and who opposes it?

Support comes from the progressive wing of Congress, a number of AI safety researchers including some employees of leading AI companies, and — according to polling cited by the sponsors — a majority of voters. Opposition comes from the White House, most of the AI industry's leadership, and the Republican majority in Congress.

In favour. Common Dreams reported a poll showing 68% of US voters back the Sanders-Casar approach of an AI pause and a superintelligence ban. Employees of major labs have endorsed it. Juan Felipe Cerón Uribe, a researcher on OpenAI's Safety Systems team, said "superintelligence could either go extremely right or extremely wrong." Swante Scholz of Google DeepMind said the current development path leads to "existential catastrophe for humanity."

Casar's argument is about who holds the capability: "Experts are warning that AI superintelligence, wielded by the wrong humans or by rogue AI, could kill countless numbers of people."

Against. President Trump was direct: "We will only encourage superintelligence. We're going to encourage it, not rein it in." He also said his administration would use the term "super intelligence" in place of "artificial intelligence" in government documents.

Inside the industry the picture is more divided than it looks. Meta's Mark Zuckerberg rejected coordinated slowdown proposals in an NBC News interview the same week. That put him at odds with other frontier-lab leaders who had voiced support for some form of coordinated slowdown. Backing a coordinated slowdown in principle is a long way from backing a criminal ban, but the industry is not speaking with one voice.

Broader public concern is rising too. PBS cited polling that 53% of Americans are "extremely" or "very" concerned about AI's environmental impact, up from 41% a year earlier — a concern tied to the build-out we covered in the AI data center power crunch.

Will the Ban Artificial Superintelligence Act pass?

No, not in this Congress. The bill has no Republican co-sponsor, no committee hearing or vote scheduled, and a president who has publicly committed to the opposite policy. PBS described it as facing "long odds in the Republican-controlled Congress."

The more useful question is what it changes without passing. Three things:

1. It sets the outer boundary of the debate. Before this, the ambitious position in US policy was licensing and transparency. A criminal ban with nuclear-style penalties makes licensing look like the moderate compromise.
2. It becomes campaign material. PBS noted the bill is part of growing positioning on AI ahead of the 2028 presidential race. Expect candidates to be asked whether they support it.
3. It gives other jurisdictions a template. The bill explicitly calls for international agreements and export controls. Its text can be borrowed by legislators elsewhere even if it dies in committee in Washington.

There is also a structural weakness the sponsors have not resolved. A US-only ban does not stop development elsewhere, and the international provisions are a directive to pursue agreements, not an agreement. Without enforceable verification between rival states, a domestic ban mainly decides where the work happens.

Frequently asked questions

What is the Ban Artificial Superintelligence Act? The Ban Artificial Superintelligence Act is a US bill introduced on September 23, 2026 by Sen. Bernie Sanders and Rep. Greg Casar. It would permanently ban superintelligent AI, pause advanced AI development until safety rules exist, and create a cabinet-level Department of Artificial Intelligence.

How does the bill define superintelligence? The bill defines superintelligence as an AI that exceeds human cognitive performance and capabilities across most domains, or that has sufficient capabilities to destroy or disempower humanity, including by overthrowing the federal government. Either condition is enough on its own.

What are the penalties under the Ban Artificial Superintelligence Act? Individuals would face up to 20 years in prison and companies would face dissolution, which the bill calls the corporate death penalty. The sponsors say these mirror existing penalties for unlawfully developing nuclear weapons.

Will the Ban Artificial Superintelligence Act become law? It is very unlikely in the current Congress. The bill has no Republican co-sponsor and no hearing scheduled, Republicans control Congress, and President Trump has said his administration will encourage superintelligence rather than restrict it.

Does the bill ban ChatGPT or Claude? Not explicitly. It targets superintelligence and pauses "advanced" AI development, and whether current models fall under that would depend on how the proposed Department of Artificial Intelligence interpreted the definitions. Ordinary API users are not the target.

Who supports the Sanders AI bill? Supporters include the bill's progressive sponsors, a number of AI safety researchers including some employees of OpenAI and Google DeepMind, and, according to a poll cited by supporters, 68% of US voters.

The verdict

The Ban Artificial Superintelligence Act will not pass, and its central definition is too vague to enforce as written. "Exceeds human cognitive performance across most domains" is not a test anyone can run, and handing that judgement to an agency that does not exist yet is not a plan.

It is still the most consequential AI bill introduced this year, because it moves the question. The argument used to be whether frontier AI should be regulated at all. After a model rated Critical for autonomous exploit development shipped to the public, the argument is now whether the right analogy is consumer software or nuclear material. The bill's best idea — writing rules against measurable behaviours like shutdown subversion — deserves to survive in whatever legislation comes next.

If you build with these models, nothing changes on Monday. What changes is the range of outcomes you should plan for. For the security side of that, read our guide to AI coding agent security.

Nobody in Washington expects this bill to become law. A year ago, nobody expected it to be written.