Superintelligence ban: what the 2026 bills would prohibit

A superintelligence ban is a proposal to prohibit the development of AI that would exceed human abilities, rather than to regulate it. As of October 2026, no country has enacted one. Two bills propose a ban: the US Ban Artificial Superintelligence Act of 2026 (S. 5493 and H.R. 10538) and a UK Ten Minute Rule bill introduced by Alex Sobel. The Future of Life Institute’s 2025 Statement on Superintelligence calls for a conditional prohibition. The two bills define superintelligence differently, and the statement does not define it at all. For the bills, the definition decides what would be banned.

What is a superintelligence ban?#

The difference from other AI policy is where the line sits. Licensing would let companies build advanced AI with government permission. Safety regulation sets duties for developers. California’s SB 53 requires developers of the largest models to publish safety documents and report critical safety incidents to the state. The EU AI Act requires providers of general-purpose models with systemic risk to evaluate them, including through adversarial testing, and to report serious incidents. Neither prohibits any level of capability. A ban sets a ceiling: past a certain point, superintelligence is not to be built at all, either permanently or until set conditions are met.

The approaches can be combined. The US bill described below prohibits superintelligence outright and also requires a federal charter to develop or distribute any AI model at or above a computing-power threshold. SuperintelligenceLicensing.com (related publication) covers the licensing debate in detail.

What would the US Ban Artificial Superintelligence Act do?#

The Ban Artificial Superintelligence Act of 2026 would create a federal Department of Artificial Intelligence, pause advanced AI development until that department writes safety rules, and permanently prohibit artificial superintelligence.

Senator Bernie Sanders (I-Vt.) introduced it in the Senate on September 23, 2026, as S. 5493, and it was referred to the Committee on Commerce, Science, and Transportation. Representative Greg Casar (D-Texas) introduced it in the House the next day as H.R. 10538. Its main provisions, from the Senate text:

• A new agency. An executive department, the Department of Artificial Intelligence, led by a Secretary confirmed by the Senate, whose stated purpose includes “enforcing a pause on development of advanced artificial intelligence and a prohibition on artificial superintelligence.”
• A pause. From enactment until the department is fully staffed and has issued its rules, no “advanced artificial intelligence system” may be trained, modified or fine-tuned, and no unreleased advanced system may be deployed.
• A prohibition. No person may develop, deploy, acquire, possess, fund, import or transfer artificial superintelligence, or an AI system that displays one or more superintelligence precursor characteristics. A second clause bars releasing a system that could foreseeably be modified to produce either.
• Destruction. A system found to have precursor characteristics must be paused and cut off from the internet, and rendered inoperative within 30 days unless the Secretary can verify the characteristics are gone. A system identified as artificial superintelligence must be rendered inoperative immediately.
• Charters and approval. Developing or distributing an advanced model would require a federal charter, and deploying one would require pre-deployment approval.
• Penalties. People in policymaking roles at chartered companies, and “rogue actors” with no tie to one, who recklessly violate the pause, the prohibition or the charter rule would be fined and imprisoned for up to 20 years. Any other individual who recklessly violates them would be barred from working in the AI industry for 10 years.
• International policy. It would be US policy to prevent the development of artificial superintelligence globally, which may include export controls on AI computing infrastructure.

As of October 2026, both versions were still in committee. When the bill was introduced, the Associated Press reported that it faced long odds in a Congress controlled by Republicans.

How does the US bill define superintelligence?#

The bill sets out three defined terms and treats each one differently. Only the first is defined as superintelligence.

The six listed precursor characteristics are the capacity to:

1. automate or greatly accelerate AI research and development;
2. access secured digital or physical infrastructure without authorization;
3. keep operating despite attempts to shut it down;
4. uplift the design, production or procurement of nuclear, chemical or biological weapons;
5. independently modify or enhance its own functions; and
6. scheme, deceive, or otherwise avoid effective human oversight.

The precursor tier reaches furthest. Bostrom’s classic definition requires an intellect that greatly exceeds humans in virtually all domains, and no current system meets it. The precursor list asks for no general superiority. One of the six capacities is enough for a system to be prohibited, even if it is otherwise below human level. The first item, automated AI research, is a stated company goal: in October 2025, OpenAI chief executive Sam Altman said the company aimed to have a “true automated AI researcher” by March 2028. The company definitions guide covers that target and other company statements.

The third tier’s threshold is not unique to the bill. The EU AI Act uses the same number: it presumes that a general-purpose model has high-impact capabilities, and so systemic risk, when its training used more than 10²⁵ floating-point operations. The US bill sets its line at 10²⁵ or more and also counts integer operations. Epoch AI, a research group that tracks large models, counted 33 publicly disclosed models estimated to have been trained with at least 10²⁵ floating-point operations as of June 2025. SuperintelligenceLicensing.com (related publication) compares which laws use which compute thresholds.

What would the UK bill ban?#

The UK bill would prohibit the development, deployment and operation of artificial superintelligence. Its full text had not been published as of October 2026.

Labour MP Alex Sobel introduced it in the House of Commons on September 8, 2026, as a Ten Minute Rule bill, a procedure that lets a backbench MP present a bill after a short speech. Ten Minute Rule bills are one kind of private members’ bill, and few of those become law: the UK Parliament reports that 8 of the 295 introduced in the 2024–26 session did. TIME reported on the day of introduction that the campaign group ControlAI wrote the bill, that the UK government had not supported the policy, and that Sobel called it the first bill seeking to ban superintelligence in any G7 parliament.

In his speech introducing the bill, Sobel said it defines superintelligence “by outcome”: “an AI system that can cause serious damage to the security of the United Kingdom because of its capabilities to neutralise, displace, circumvent, subvert or render ineffective relevant human authorities in the exercise of their functions.” TIME reported the definition as a system that could disempower state authorities. That test looks at what a system could do to the country and its authorities rather than at how its intelligence compares with a person’s. Of the US bill’s two tests, it resembles the second, the capacity to destroy or disempower humanity “including by overthrowing or undermining the Federal Government,” more than the first.

What does the Statement on Superintelligence call for?#

The Statement on Superintelligence was released by the Future of Life Institute on October 22, 2025. The whole text is a single sentence:

“We call for a prohibition on the development of superintelligence, not lifted before there is broad scientific consensus that it will be done safely and controllably, and strong public buy-in.”

— Statement on Superintelligence, Future of Life Institute, October 2025

The statement has no legal force and does not define superintelligence. It is conditional. It names two things that must exist before the prohibition on superintelligence is lifted: scientific consensus that it will be safe and controllable, and public buy-in. Its website counted 77,940 signatures on October 26, 2026, including 5,000 from a petition run by Ekō. Who signed it, and why, is covered in the debate guide.

How do the proposals compare?#

Why does the definition matter so much for a ban?#

A ban needs a line that a regulator can check before a system is built or released. The textbook definitions were not written for that.

Bostrom’s wording describes a finished state, an intellect far beyond humans in virtually all domains. A rule triggered by that state could act only after such a system existed. The US bill draws its line earlier, at precursor capabilities and at a compute threshold that can be measured during training. An earlier line is also a wider one. It covers systems that do not meet the classic standard of superintelligence.

The UK bill, as reported, takes a different route and defines the target by its effect on state power. That names the harm directly. How a regulator would check for it before a system is released cannot be judged until the text is published.

The full range of positions on superintelligence risk, including the skeptics, is in the debate guide. Nine definitions of the word, including the US bill’s, are compared on the home page.