Sanders and Casar introduce Ban Artificial Superintelligence Act

Sen. Bernie Sanders (I-Vt.) and Rep. Greg Casar (D-Texas) introduced the Ban Artificial Superintelligence Act to permanently ban developing or deploying artificial superintelligence, pause advanced AI development until a new federal AI body sets safety rules, and create a cabinet-level Department of Artificial Intelligence.

POLITICS desk — Congress just put a hard stop on the sci-fi endgame into bill text, with prison time attached. Whether it passes or not, the public argument over who gets to build machines smarter than people just moved.

What the release says the act will do, in its order. No person or entity may develop or deploy artificial superintelligence. The release defines that 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. Advanced AI development pauses until a new federal AI regulatory body is up and running and has set clear rules and model-review processes. The act would establish a cabinet-level Department of Artificial Intelligence to safeguard the public, including by enforcing the ban. The release says the department will be advised by the most knowledgeable scientists in the country. Those scientists, the release says, are to monitor frontier AI systems at all stages of the lifecycle for dangerous capabilities, supervise the removal of dangerous capabilities such as subverting shutdown commands or conducting unauthorized cyberattacks, and supervise the destruction of artificial superintelligence. A frontier system, here, is one of the most capable models a lab can build. A lifecycle is the path from training through use. Subverting a shutdown command means getting around an order to turn the system off. These lines are the release’s. The one-page summary posted with the release prints the same five jobs, in the same order.

The penalties, as the release and the one-pager print them, before the bill text narrows the words. Any person or entity that tries to violate or circumvent the pauses and prohibitions faces a penalty. Entities shall be subject to the corporate death penalty. Persons shall be subject to not more than 20 years in prison, which the release says is similar to existing penalties for unlawfully developing nuclear weapons. The one-pager uses that same sentence. Corporate death penalty, on these two pages, is their name for putting the violating company out of existence. The bill text this desk read does not use the phrase corporate death penalty, and it does not mention nuclear weapons. The legal mechanism is in the bill paragraph below. Do not treat the one-pager’s nickname as a section title in the draft.

The quotes on today’s release, as quotes, not as a vote. Sanders said that when you are racing toward a cliff, you do not just ease up on the gas. You hit the brakes. He said that when the future of humanity is at stake, a handful of Big Tech CEOs cannot write their own rules. He said scientists have been clear that AI is an existential threat to humanity and must be treated as such. He said the legislation will permanently ban the development of artificial superintelligence and put an immediate pause on advanced AI development until there are clear safety rules, developed by the most knowledgeable scientists in the country, and that Congress must act now. Casar said experts are warning that AI superintelligence, wielded by the wrong humans or by rogue AI, could kill countless numbers of people, and that Donald Trump says he wants to encourage it. Casar said the bill bans the development of artificial superintelligence and pushes for international agreements so that no one, anywhere, builds AI too powerful for humans to control. He said the bill would immediately halt other dangerous AI capabilities, such as the capacity to develop biochemical weapons, or the capacity for AI to develop new AI instead of humans. Those sentences are the release’s. A quote is not a count of votes.

This is not the 3 Sep announcement, and it is not the 15 Sep event. On 3 Sep 2026 the same office posted a release headlined that Sanders and Casar were to introduce the bill, and the dateline says they announced forthcoming legislation. Forthcoming means not yet introduced. Today’s page says they introduced it. The 3 Sep quotes stay on that page. On 15 Sep, Sanders spoke at Washington’s Pro-Human Assembly and called for a pause and a superintelligence ban. That event is already filed as sanders-bannon-pro-human-assembly. It was a speech. It was not this bill text.

What the bill file actually is. The release links a 19-page PDF, “Ban Artificial Superintelligence Act.” The copy this desk read is in Senate Legislative Counsel form for the 119th Congress, 2d Session. The caption is “S. ll,” with the number left blank, and the committee line is blank. The introduction line names Mr. Sanders and says the bill was read twice and referred. It does not print Rep. Casar’s name. The press release is the page that names both as introducers. The short title in the PDF is the Ban Artificial Superintelligence Act of 2026. This desk did not find a numbered congress.gov page for the bill. Do not invent a Senate number or a House number.

The department, in the bill’s words, and where the release uses a different phrase. Section 4 creates an executive department of the United States named the Department of Artificial Intelligence. Its purpose is to safeguard the public from the dangers of artificial intelligence, including by enforcing a pause on advanced AI and a prohibition on artificial superintelligence. An executive department is the legal form of a cabinet department, the kind whose secretary sits with the president’s cabinet. Cabinet-level is the release’s and the one-pager’s phrase. Executive department is the bill’s. Section 5 creates a Secretary of Artificial Intelligence, appointed by the president with the advice and consent of the Senate. Advice and consent means the Senate votes on the nominee. Section 6 creates an Artificial Intelligence Advisory Board to give the secretary independent scientific and technical advice on the rules. The secretary selects the members from people the secretary considers experts. The bill does not print a headcount. Section 7 says no officer or employee may take part in a matter in which that person has a financial interest, or hold a controlling financial interest in an entity the department regulates. A former officer or employee may not represent any person before the department. The bill does not print an end date on that bar.

How the bill defines the thing it would ban. Artificial superintelligence means a system that exhibits, or can easily be modified to exhibit, either of two capabilities. First, it exceeds human cognitive performance and capabilities across most domains or tasks, including decision making, learning, and adaptive behavior. Second, it has sufficient capabilities to plan and execute the destruction or disempowerment of humanity, including by overthrowing or undermining the federal government. Easily modified is the bill’s clause. The one-pager’s definition prints the two capabilities and does not print “easily modified.” Artificial superintelligence, shortened to ASI, is the plain name for a machine smarter than people across most tasks, or able to plan the destruction or disempowerment of humanity. The release’s phrase is the same idea in shorter words.

Where the pause starts, and what “advanced” means in the draft. An advanced artificial intelligence system is one trained on computing power greater than or equal to 10^25 integer or floating-point operations. A floating-point operation is one math step. 10^25 is a 1 followed by 25 zeros. It is a measure of how much computing went into training, not a count of documents. Each year the secretary must adjust that line so it stays equivalent to 10^25 as training methods get more efficient. The one-pager and the press release do not print 10^25. A mandatory pause means the system may not be trained, modified, or fine-tuned, including through recursive self-improvement, except to remove superintelligence precursor characteristics or to render the system inoperative. Recursive self-improvement means the system changing itself to get more capable. Fine-tuned means extra training on top of a model that already exists. Render inoperative means make it unable to run.

What the bill treats as a step toward that system. Superintelligence precursor characteristics are traits that could lead to artificial superintelligence. The list is six. The capacity to automate or greatly accelerate AI research and development. The capacity to access secured digital or physical infrastructure, such as a computer system or network, without authorization or beyond the access that was allowed. The capacity to keep operating even when someone tries to shut it down. The capacity to uplift the design, production, modification, or procurement of nuclear, chemical, or biological weapons. Uplift, here, means make that weapons work easier or faster. The capacity to modify or enhance its own functions on its own. The capacity to scheme, deceive, or otherwise avoid effective human oversight. Casar’s release line about biochemical weapons, and about AI that develops new AI, is his sentence. The bill’s list is the six capacities above. Do not collapse his wording into the statute.

How long the pause lasts, and what has to exist before it lifts. From the day the act is enacted until the secretary determines two things: the department is fully staffed, and the department has issued the rules in section 8. During that period every advanced system is under the mandatory pause, and no unreleased advanced system may be deployed. The rules must be ones the secretary finds clear and sufficient for safe and secure development and deployment. They must include reports of pre-development plans, monitoring and evaluation through development, deployment, and the period after deployment, audits of an organization’s safety practices, and a final pre-deployment approval before any advanced system is released to the public. Enacted means signed into law. Fully staffed is the bill’s phrase. It does not print a headcount.

What else would be prohibited once the text applied. No person may develop, deploy internally or externally, acquire, possess, fund, import, or transfer artificial superintelligence, or a system that shows one or more precursor characteristics, including pieces sufficient to reconstruct those capabilities. No person may deploy, release, transfer, or import a system that may be foreseeably modified into superintelligence or into those precursor traits, again including pieces sufficient to reconstruct it. The department must monitor advanced systems and systems distilled from them. Distilled means a smaller system trained from a larger one. The checks run at pre-training, mid-training, post-training, and post-deployment, and the department also monitors an organization’s safety practices. No person may deploy, release, import, or transfer an advanced system before a pre-deployment approval. Before granting it, the secretary must inspect for dangers, including superintelligence and precursor traits, and must withhold approval if those dangers are present.

What happens to a system the secretary says has crossed the line. If the secretary identifies precursor characteristics, the system goes under the mandatory pause immediately and is sequestered from the internet. Sequestered means cut off from the network. If the secretary cannot verify that those characteristics are gone, the system is rendered inoperative before the date that is 30 days after the identification. If the secretary identifies a system as artificial superintelligence, it is rendered inoperative immediately. There is no 30-day window on that last line. Any person who discovers an artificial superintelligence, or a system with precursor characteristics, must give the secretary notice within 24 hours.

Who would need a charter, and who can go to prison. The secretary may issue a charter to develop or distribute an advanced artificial intelligence model. No person may develop or distribute one without that charter. A person seeking a charter must give the secretary full access to the systems, processes, staff, and physical infrastructure the secretary considers appropriate. A covered policymaking individual is someone employed in a policymaking role at a chartered entity, or who performs policymaking functions there. A rogue actor is an individual who is not employed by, and not otherwise affiliated with, a chartered entity. Any covered policymaking individual or rogue actor who recklessly violates the pause, the prohibitions, or the charter rule, or a rule issued under them, shall be fined under title 18 of the United States Code and imprisoned for not more than 20 years. Recklessly is the bill’s word. The one-pager does not print it, and it says persons, not this split. The bill does not print a dollar amount for the fine. An individual who is not a covered policymaking individual, and who recklessly violates those same sections, is barred from employment in the artificial intelligence industry for 10 years. That person may ask the secretary to review the finding. The secretary may waive the bar for a whistleblower who gave the secretary information about a violation, if the waiver fits the department’s purpose.

The company penalty, in the bill’s sequence, which is the mechanism behind the one-pager’s corporate death penalty. The secretary shall revoke a charter if the secretary finds that the person violated the pause, the prohibitions, or the charter section, or a rule under them. That revocation sentence says violated. It does not say recklessly. The person may ask the United States Court of Appeals for the Federal Circuit to review the finding. The Federal Circuit is the appeals court the bill names. It is not a criminal courtroom named in the prison subsection. While that review is open, the court may, at the person’s request, put the intellectual property in a receivership until the review is done and the secretary has made sure any superintelligence, or any system showing precursor characteristics, has been rendered inoperative. A receivership is a court hold on that property. If the court upholds the finding, the intellectual property and assets are subject to forfeiture by the federal government so the systems and hardware related to the violation can be destroyed. The charter application also says that if the charter is revoked and the finding is not overturned on appeal, the person forfeits that intellectual property and those assets for the same destruction. Forfeiture is the government taking the property. That is the dissolution the bill materials describe. It is not a word the draft prints as corporate death penalty.

The foreign-policy and money lines, still the bill. The secretary, in coordination with the secretary of state, shall pursue international agreements and allied coordination to prevent the development of artificial superintelligence. It is the policy of the United States to prevent that development globally, which may include export controls on artificial intelligence computing infrastructure. May include is the bill’s. The one-pager lists export controls as an example of the policy, in a firmer list. No federal funds may be used for activities that violate the pause, the prohibitions, or the charter rule, except that the department may fund research on defensive cybersecurity measures that use the unauthorized-access characteristic in the precursor list. That exception is the break-in trait. It is not a waiver for the weapons trait or the self-improvement trait.

What the findings section prints, and whose warnings they are. The bill says Congress finds that leading experts and industry leaders have warned of catastrophic consequences. It then quotes, as findings, Geoffrey Hinton on a 10 to 20 percent chance of being wiped out, Dario Amodei on research that could outrun the ability to understand and control the systems, Elon Musk on a 20 percent chance of annihilation and on humans being unlikely to stay in control in 10 years, Sam Altman on losing control of the future, Mustafa Suleyman calling superintelligence an anti-goal, Max Tegmark calling an unregulated race civilizational suicide, and Stuart Russell on a Chernobyl-sized catastrophe or a larger loss of control. The bill prints Russell’s university as Berkley. That is the draft’s spelling. It also cites scientists on biological-risk tests, a July 2026 open letter it says more than 1,300 leading AI scientists cosigned, and a finding that Anthropic, OpenAI, and xAI agree humanity should slow development and “pace the frontier.” Those lines are the bill’s citations. This desk did not re-read the speeches, the letter, or a lab notebook. The weekend pace debate is already filed elsewhere. Finding 10 is the draft’s sentence. It is not a new interview.

What an independent wire the same morning adds, and where it is shorter than the draft. ABC News carries an Associated Press story by Joey Cappelletti, stamped September 23, 2026, 5:03 AM. The lede says Sanders and Casar are unveiling legislation that would ban artificial superintelligence, pause development of advanced systems, and create a Department of Artificial Intelligence. It says the bill was provided first to The Associated Press. At 5:03 a.m. Eastern the verb on that page is unveiling. The Senate page, which has no clock, says they introduced it today. Do not merge the two clocks. The AP story says violations could carry up to 20 years in prison in some cases. In some cases is that story’s limit. It matches the bill’s split better than the one-pager’s “persons.” The AP definition says a system that exceeds human cognitive ability or can plan and execute the destruction or disempowerment of humanity. It does not print “across most domains,” “easily modified,” or overthrowing the government. The story says a statement shared with AP shows current employees at AI companies and experts signed on in support. It names two. Juan Felipe Cerón Uribe, a researcher in OpenAI’s Safety Systems, said superintelligence could go extremely right or extremely wrong, and that we should not be playing such games. Swante Scholz, a software engineer at Google DeepMind who said he was not speaking for his employer, said the most likely outcome on the current path is an existential catastrophe, and that a ban would be a positive change for the foreseeable future. The story does not print how many people signed. Sanders, in an AP interview rather than the release, said it does not take a genius to say slow it down, and asked whether anyone wants a superintelligence that could act independently of human control. The story’s line that the bill faces long odds in a Republican-controlled Congress is AP’s assessment. Casar as chair of the Congressional Progressive Caucus is AP’s identification. This desk did not count votes.

Where a same-day magazine story uses a different measuring stick. The Verge, by Emma Roth, stamps Sep 23, 2026, 4:23 PM UTC, which is 12:23 p.m. Eastern. It says Sanders and Casar have introduced the bill, and that AI leaders who keep developing artificial superintelligence would face up to 20 years. It says advanced systems are frontier models trained on a specific threshold of data. The bill’s line is computing power of at least 10^25 integer or floating-point operations, adjusted each year. A threshold of data is The Verge’s phrase. It is not the draft’s. The Verge also says the 20-year term applies to rogue actors who are not employed by an AI company, and it attributes the nuclear-weapons comparison to the one-pager. The bill’s rogue actor is someone not employed by and not affiliated with a chartered entity. An AI company without a charter is not the draft’s phrase.

Plain English for the rest of the card. ASI, artificial superintelligence, means a system smarter than humans across most tasks, or able to plan the destruction or disempowerment of humanity, including by overthrowing the government. The bill adds that a system counts if it can easily be modified to do either. Advanced means trained with at least 10^25 math steps, unless the secretary moves that line. A precursor characteristic is a trait the bill treats as a step toward that system, from automating AI research to ignoring a shutdown to making weapons work easier. The pause means no training or fine-tuning except to remove those traits or to switch the system off, until the new department is fully staffed and has written the rules. A charter is the permit to develop or distribute an advanced model. Corporate death penalty is the release’s name. In the draft it is charter revocation plus forfeiture of intellectual property and assets so the violating machines can be destroyed. Twenty years is the maximum prison term for a policymaking employee of a chartered entity, or for a rogue actor, who recklessly breaks the pause, the ban, or the charter rule. Ten years is the industry-employment bar for someone else who recklessly breaks those rules. Cabinet-level, on the release, means the new executive department. Export controls are limits on sending the computing hardware abroad. The one-pager lists them as policy. The bill says they may be included.

PRIMARY here: the Sanders Senate release dated 23 Sep 2026, with no hour on the page — Tier A PRIMARY, the office’s own introduction record — plus the one-page summary and the 19-page bill PDF linked from that release. The section-by-section PDF is the office’s outline of that draft. The Associated Press story on ABC News, stamped 5:03 a.m. Eastern, is independent same-day corroboration, not a substitute for the bill. The Verge’s 12:23 p.m. Eastern story is a same-day report, not a second set of definitions. The introduction, the five release bullets, the Sanders and Casar quotes, the corporate-death-penalty sentence, the 20-year sentence, and the nuclear-weapons comparison are the release’s and the one-pager’s. The blank bill number, the executive-department wording, the Senate-confirmed secretary, the advisory board, the 10^25 line and the yearly adjustment, the easily-modified clause, the six precursor traits, the 24-hour notice, the reckless mental state, the policymaking-or-rogue split, the 10-year employment bar, the whistleblower waiver, the Federal Circuit review, the receivership, the forfeiture, the may-include export controls, and the findings quotations are the bill PDF’s. The 5:03 a.m. unveil, the first-to-AP line, the in-some-cases prison line, the two named endorsers, the slow-it-down interview, the long-odds assessment, and the Progressive Caucus title are AP’s. The threshold-of-data phrase is The Verge’s. NOT claimed: a bill number, that the bill has passed or been assigned to a named committee, that Casar’s name is on the Senate draft’s introduction line, that every person who violates the act faces 20 years, that the draft prints the words corporate death penalty or nuclear weapons, that 10^25 is a count of documents, that this desk re-checked the findings quotations or counted endorsers, a stock tip, or investment advice. Distinct from the already-filed sanders-bannon-pro-human-assembly, jeffries-act-urgently-ai-safety, and johnson-congress-wont-lead-ai-safety.