The US Bill That Would Ban "Superintelligent" AI, Explained
Sen. Bernie Sanders and Rep. Greg Casar formally unveiled the Ban Artificial Superintelligence Act on September 23, 2026. This explainer covers what the bill would ban, how its compute pause and new Department of AI would work, the penalties, and why it is unlikely to pass soon.
⚡ Quick facts
- Sponsors: Sen. Bernie Sanders (I-Vt.) and Rep. Greg Casar (D-Texas); formally unveiled September 23, 2026
- The ban: permanently bars developing or deploying artificial superintelligence in the US
- The pause: halts work on AI systems trained with 10^25 or more operations until a new regulator sets rules
- Penalties: up to 20 years in prison for individuals; companies lose their charter and assets
- Odds: long: Republicans control Congress and the administration favours accelerating AI
What the bill would do
The Ban Artificial Superintelligence Act was trailed in early September 2026 and formally unveiled by Sen. Bernie Sanders and Rep. Greg Casar on September 23, 2026 (The Next Web). It has four main parts: a permanent ban on superintelligence, an immediate pause on the most advanced AI training, a new cabinet-level Department of Artificial Intelligence, and a push for international agreements.
How it defines "superintelligence"
The bill covers AI that "exceeds human cognitive performance and capabilities across most domains" (Common Dreams), or a system capable of planning and carrying out the destruction or disempowerment of humanity. Developing or deploying such a system would be banned outright, with no licensing route. Casar also said the bill targets recursive self-improvement and bars using AI to design nuclear or biological weapons.
The pause and the charter system
Work on advanced AI systems trained with 10^25 or more operations would pause immediately. The pause would end only once the new Department of AI is fully staffed and has set rules. After that, any company building or selling advanced AI would need a charter from the department. The department would also monitor frontier systems, order dangerous capabilities removed, and oversee the destruction of any superintelligence. A system's ability to change its own functions is listed as a warning sign.
Penalties and protections
Individuals who break the ban would face up to 20 years in prison. Companies would lose their charter and hand their intellectual property and assets to the federal government, which supporters call the "corporate death penalty". Employees who report superintelligence risks would get whistleblower protection. The bill also directs the US to pursue international agreements, coordination with allies, and export controls on AI compute, so that superintelligence cannot simply be built elsewhere.
Will it pass?
Almost certainly not in this Congress. Republicans control both chambers, and other AI bills are already waiting with no vote scheduled. The administration's line is to encourage AI, not pause it. Sanders and Casar want the bill to be a top Democratic priority if Democrats win the House and Senate in the November 2026 midterms. Some AI-lab employees have publicly backed it, speaking for themselves. Law firm Akin Gump is tracking it as a bill that has been introduced. Treat it as a marker of where the debate is heading, not as imminent law.
Our take: what the 10^25 line means in practice
The 10^25 figure is not new. The EU AI Act uses the same threshold, 10^25 floating-point operations of training compute, to presume that a general-purpose model carries "systemic risk" (EU AI Act, Article 51). The difference is the consequence. In the EU, crossing the line brings extra testing and reporting duties. Under this bill, crossing it would stop training altogether until a regulator exists. If you follow AI policy, watch three things:
- Cosponsors. No cosponsor list was released at launch. Any Republican support would change its odds.
- The compute threshold. A fixed 10^25 line catches more models every year as training gets cheaper. Watch for amendments that index it.
- The definition. "Most domains" is hard to test. How a regulator would measure it decides whether the ban is workable.
Sources
- The Next Web: Sanders bill would ban superintelligence and create a Department of AI
- Common Dreams: Sanders and Casar unveil superintelligence ban
- Akin Gump AI law tracker: Sanders and Casar introduce bill
- EU AI Act, Article 51: classification of systemic-risk models
Frequently asked questions
What is the Ban Artificial Superintelligence Act?
A US bill from Sen. Bernie Sanders and Rep. Greg Casar, formally unveiled on September 23, 2026. It would permanently ban superintelligent AI, pause the most advanced AI training, and create a cabinet-level Department of Artificial Intelligence.
How does the bill define superintelligence?
AI that exceeds human cognitive performance and capabilities across most domains, or a system able to plan and carry out the destruction or disempowerment of humanity.
Which AI models would be paused?
Development of advanced AI systems trained with 10^25 or more operations would pause until the new Department of AI is staffed and has set rules. After that, companies would need a charter from the department.
What are the penalties?
Up to 20 years in prison for individuals. Companies would lose their charter and hand their intellectual property and assets to the federal government.
Will the bill become law?
It faces long odds. Republicans control Congress and the administration favours accelerating AI. The sponsors want it to become a Democratic priority after the November 2026 midterms.
How does it compare with the EU AI Act?
Both use a 10^25 compute threshold. The EU AI Act uses it to presume systemic risk and add testing and reporting duties. This bill would use it to pause training until a new US regulator sets rules.