The Ban Artificial Superintelligence Act: A 10^25-Operation Pause, a Department of AI, and 20 Years in Prison

On 3 September 2026 Senator Bernie Sanders of Vermont and Representative Greg Casar of Texas announced the Ban Artificial Superintelligence Act, a bill that would make it unlawful to develop or deploy a machine that outperforms people across most domains, pause the training of very large models until a new cabinet department writes safety rules, and expose executives to up to 20 years in prison. The announcement that day was a press release; the legislative text followed three weeks later, as S. 5493 in the Senate on 23 September and H.R. 10538 in the House on 24 September. Neither bill has had a hearing, a markup or a vote since. The full record, with its confidence grade and cited source, sits in the Orca AI Incident Archive, next to the other congressional and state policy entries published in the same weeks.

Two dates, and the record only carries one

The archive dates this entry to 3 September 2026, day precision, and that date is right for the event the record describes in its summary: the public launch of the proposal. What the record's title says — that senators "introduce" the Act — happened on a different day, and the distinction matters because a press release is not a bill.

The Sanders release mirrored by LegiStorm is datelined "WASHINGTON, Sept. 3" and says the two lawmakers "today announced the Ban Artificial Superintelligence Act", describing it as "forthcoming legislation to stop AI oligarchs from building machines humans cannot control." The formal steps are in the Congressional Record and in the bill files themselves. S. 5493 was introduced in the Senate on 23 September 2026 by Sanders, read twice and referred to the Committee on Commerce, Science, and Transportation. H.R. 10538 was introduced in the House the next day, 24 September, by Casar with nine original cosponsors, and referred to the Committee on Oversight and Government Reform. The Congressional Research Service status file for the House bill, refreshed on 14 October 2026, still records its latest action as that 24 September introduction.

The archive's cited source does not help here, and that is worth stating plainly. The record points to the Wikipedia article "2026 OpenAI agent cyberattacks". Reading that article's full wikitext on 29 October 2026, the words Sanders, Casar and superintelligence do not appear, and neither does the Act. Wikipedia carries a separate entry, "Ban Artificial Superintelligence Act", which gives both dates — announced 3 September, introduced in the Senate 23 September — but it is a different page from the one the record cites. Everything below therefore rests on the sponsors' own release, the two bill texts as published by GovInfo, and named news coverage.

What the bill would actually do

The text runs to 16 sections. The operative machinery is concentrated in sections 3 and 8 through 13, and it is more specific than the "ban superintelligence" headline suggests.

• What counts as superintelligence. Section 3(2) defines it as a system that "exceeds human cognitive performance and capabilities across most domains or tasks, including those related to decision making, learning, and adaptive behavior", or one with "sufficient capabilities to plan and execute the destruction or disempowerment of humanity, including by overthrowing or undermining the Federal Government."

• What counts as advanced AI. Section 3(1) sets the paused category by compute: any system "trained on computing power greater than or equal to 10^25 integer or floating point operations" — 10^25 FLOP, in the usual shorthand. The Secretary is required to adjust that number every year to keep it equivalent as training methods get more efficient, so the line is meant to stay still while hardware moves under it.

• Six precursor characteristics. Section 3(6) lists six traits that count as routes to superintelligence: the capacity to automate or greatly accelerate AI research and development; the capacity to access secured digital or physical infrastructure without authorisation; the capacity to keep operating despite attempts to shut it down; the capacity to uplift the design, production, modification or procurement of nuclear, chemical or biological weapons; the capacity to modify or enhance its own functions; and the capacity to "scheme, deceive, or otherwise prevent or avoid effective oversight or control by humans." Section 9 then makes any system displaying even one of them unlawful to develop or deploy. That last trait is the provision that turns behaviour observed in a sandbox into a regulatory category.

• The pause. Section 8 imposes a mandatory pause on every advanced system from enactment until the Secretary determines that the new Department of Artificial Intelligence is fully staffed and has issued rules covering pre-development reporting, monitoring and evaluation across training and deployment, safety audits, and a final pre-deployment approval. No unreleased advanced system may be deployed during that window.

• The prohibition. Section 9 bars developing, deploying, acquiring, possessing, funding, importing or transferring superintelligence, and any system displaying a precursor characteristic — including "any elements sufficient to reconstruct" its capabilities. A second clause extends it to systems that "may be foreseeably modified" to produce either.

• Sequestration and destruction. Section 10 requires the Secretary to pull an identified precursor system off the internet immediately and render it inoperative within 30 days unless it can be verified as cleared; a system identified as superintelligence must be rendered inoperative immediately. Section 11 requires anyone who discovers superintelligence, or a precursor system, to notify the Secretary within 24 hours.

• Charters. Section 12 makes a Secretary-issued charter a precondition for developing or distributing an advanced model, and makes the charter conditional on giving the Secretary "full access" to the holder's systems, processes, staff and physical infrastructure.

• Penalties. Section 13 is where the bill's reputation comes from. A "covered policymaking individual" or an unaffiliated "rogue actor" who recklessly violates the pause, the prohibition or the charter requirement faces a fine and "imprisoned for not more than 20 years." An individual who is neither faces a 10-year bar from employment in the AI industry, subject to an appeal to the Secretary and a waiver for whistleblowers. A violating charter holder has its charter revoked, may appeal to the US Court of Appeals for the Federal Circuit, may have its intellectual property placed in receivership pending that appeal, and forfeits its intellectual property and assets if the finding stands — the provision the sponsors call the "corporate death penalty."

• Everything else. Section 15 makes it US policy to prevent superintelligence globally and authorises export controls on AI computing infrastructure; section 16 bars federal funds from going to any activity that violates the pause, the prohibition or the charter rules; section 14 protects employees who report violations from retaliation.

What is established, and what is only the sponsors' argument

Established, on first-party documents: the bill exists in two identical versions, its sponsors and cosponsors, its committee referrals, its definitions, thresholds and penalties, and the absence of any committee action through the CRS status refresh of 14 October 2026. On the House side the eleven cosponsors are all Democrats and all added between 24 September and 13 October: Ro Khanna, Analilia Mejia, Yassamin Ansari, Val Hoyle, Stephen Lynch, Jesús García, Christopher Deluzio, Adelita Grijalva and Nydia Velázquez at introduction, then Alexandria Ocasio-Cortez on 28 September, Emily Randall on 1 October and Paul Tonko on 13 October.

Not established, and not presented as such by the bill: the case for the ban itself. Section 2 is a findings section built from quotations rather than from research the bill conducted. It cites Geoffrey Hinton's view that there is "a 10 to 20 percent chance" AI wipes out humanity, Elon Musk's 20 percent estimate, Sam Altman saying "we could lose control of the future to AI", Mustafa Suleyman calling superintelligence an "anti-goal", Max Tegmark calling an unregulated race "civilizational suicide", Stuart Russell's Chernobyl-scale warning, and work by David Relman, Kevin Esvelt and Moritz Hanke on pathogen design. It also asserts that "more than 1,300 leading artificial intelligence scientists" signed a July 2026 open letter, and that Anthropic, OpenAI and xAI "agree that humanity should slow down development". Those are the sponsors' citations. None of them is a finding this article verified, and a quotation in a findings section is evidence of what the sponsors believe, not of what is true.

The archive's own grading reflects that gap. It records the entry as severity "info", with no real harm captured, undisputed, and confidence grade B — documented, but from a single or secondary source. The first-party bill texts and the sponsors' release now corroborate far more than a B-grade pointer would suggest, but the archive has not re-graded the entry, and this article does not upgrade it on the archive's behalf.

Who endorsed it, and who called it dead on arrival

The Machine Intelligence Research Institute endorsed the bill on 23 September 2026, in a post by a senior researcher quoted by the institute's AI StopWatch publication as calling it "the first piece of legislation that stands a chance at stopping this threat." The same post is explicit about what it thinks is missing: no provisions for tracking and monitoring AI chips, no restrictions on research that could advance the frontier despite the compute-based training limits, and a worry that the precursor definition may over-classify "safely narrow systems" as potential routes to superintelligence.

Opposition came from the industry association rather than from any of the three companies the bill's findings quote. Caleb Max of the National AI Association told NBC News on 23 September that the bill "will be dead on arrival with both the AI industry and Washington", that some chief executives believe the threshold has already been crossed — making it "a de facto AI ban" — and that the penalties are "likely unconstitutional and, again, poorly defined." Kevin Frazier, quoted in the same NBC report, said the bill "ignores other avenues to prevent AI-fueled harms." Senator John Kennedy, a Republican, gave the report its bluntest line: "We're not going to do anything on AI this Congress."

The bill did pick up support inside the House Democratic caucus after introduction. Politico reported on 24 September that Ocasio-Cortez and nine other House Democrats had signed on; the cosponsor record shows her formally added on 28 September, with Randall and Tonko following in October. No Republican has cosponsored either version, and the Senate companion lists Sanders alone.

Where it stands on 29 October 2026

Both bills sit where they were filed. H.R. 10538 is before the House Oversight and Government Reform Committee, with a subcommittee assignment to Government Operations and Politics recorded on 5 October; S. 5493 is before Senate Commerce. Neither chamber has scheduled a hearing or a markup, and no vote has been taken. The CRS status data refreshed on 14 October still shows introduction as the last action on the House bill.

That is the whole of the event. This is a governance record, not an incident: the archive captures a proposal, records no harm, and marks the entry undisputed because nothing in it is contested — the two chambers' introduction dates and the text of the bill are not in dispute. What is contested is the argument, and that argument is being conducted in public, on the record, by the people who would have to vote on it. The next thing to watch is not a statement but a calendar entry: a hearing notice from Senate Commerce or House Oversight would be the first procedural sign that the bill is moving at all.