Superintelligence Ban: The Surprising Words It Never Uses

Superintelligence ban advocacy got its biggest mainstream platform of the year on 9 September 2026, when TechCrunch published an Equity podcast episode titled “Superintelligence is coming. Should we let it?” The guest was Connor Leahy, founder of EleutherAI and of the AI safety startup Conjecture, and now the US lead at ControlAI — a non-profit whose position is not that artificial intelligence should be aligned, contained or supervised, but that superintelligent AI should be prohibited outright, everywhere, by law.

A superintelligence ban is a genuinely radical ask, and it deserves to be read closely rather than dismissed. So we read it. Not the episode — you cannot read the episode, which turns out to be part of the story — but everything ControlAI has published in writing. Every page on its campaign site, plus the 83-page policy document those pages point to.

The superintelligence ban campaign’s site comes to 34,114 words. Across all of them, the word “extinction” appears 101 times. The word “FLOP” — the unit in which every AI law currently on the books draws its line — appears zero times. “Verify” appears 5 times. “Enforce” appears 7. “Inspect” appears 0.

The numbers do exist. They are in A Narrow Path, a 23,947-word plan published in April 2025, which specifies a training threshold of 10^25 floating-point operations, a data-centre licensing threshold of 10^17 FLOP/s, and three classes of licence. That document says “licence” at almost exactly the rate the campaign site says “extinction”. Two documents, one organisation, and the arithmetic sits in the one almost nobody reads.

Table of contents

The Superintelligence Ban Case Arrived as 41 Minutes of Audio
What the Superintelligence Ban Campaign Has Actually Published
The Superintelligence Ban Site Says “Extinction” 101 Times and “FLOP” Zero
The Superintelligence Ban Needs a Definition. ControlAI Publishes Five
A Superintelligence Ban Defined by Outcome Is a Choice, Not an Oversight
The Superintelligence Ban Scoreboard Disagrees With Itself
The Superintelligence Ban Site Has Four Pages That Load Empty
No Superintelligence Ban Has Been Introduced Anywhere Yet
What a Superintelligence Ban Would Mean for Businesses Using AI
What to Ask Before You Take a Side on the Superintelligence Ban
The Superintelligence Ban Question the Audio Never Had to Answer
References

The Superintelligence Ban Case Arrived as 41 Minutes of Audio

The TechCrunch item is not an article. It is a video post — a different content type on the same site, which is why it never appeared in the publication’s main article feed on the day it ran.

What the page actually contains

The episode’s metadata gives its duration as PT41M44S: 41 minutes and 44 seconds, or 2,504 seconds. The page’s own written content runs to 171 words, of which the human-readable summary is 152. TechCrunch’s structured data declares a wordCount of 198.

The word “transcript” appears on that page zero times. It appears zero times on the Equity podcast’s index page too. There is no published text of the superintelligence ban case Leahy actually argued.

16.5 seconds of audio per published word

Divide 2,504 seconds by the 152-word summary and you get 16.5 seconds of speech for every word a reader can quote, search, or check. At a conversational 150 words per minute, roughly 6,200 spoken words exist somewhere in that file, and about 2.5% of them made it into text.

This matters more than it would for a product launch. The superintelligence ban argument is a claim about evidence and mechanism, and the only durable version of it on that page is a summary that names neither. It is a format problem, not a bad-faith one — but it is the reason we went to the written superintelligence ban material instead.

Artefact	Format	Size	Searchable text?
Equity episode	Audio/video	41m 44s (2,504 s)	No transcript
Episode summary	Prose	152 words	Yes
Page content total	Prose + player	171 words	Yes
ControlAI campaign site	23 pages	34,114 words	Yes
A Narrow Path	PDF, 83 pages	23,947 words	Yes

What the Superintelligence Ban Campaign Has Actually Published

The written superintelligence ban splits cleanly into two layers, and the split is the whole story.

The campaign layer

The site at controlai.org — the domain controlai.com redirects to it — carries 23 pages with substantive text. The blog index is the largest at 8,397 words, then a media page at 4,776 and a quotes page at 4,438. The policy pages are smaller: 1,237 words for the policy proposals, 1,084 for the Direct Institutional Plan, 263 for the draft UK bill.

The plan layer

Underneath sits A Narrow Path, written by Andrea Miotti, Tolga Bilge, Dave Kasten and James Newport, dated April 2025 and marked version 1.2. It runs 83 pages and 23,947 words. Leahy is not an author; he is thanked in the acknowledgements, alongside Anthony Aguirre and Max Tegmark.

That document is where the mechanism lives. It proposes three licences: a Training Licence for any run above 10^25 FLOP, a Compute Licence for data centres above 10^17 FLOP/s with know-your-customer rules and physical GPU tracking, and an Application Licence. It proposes an international body it calls GUARD, and it models breakout time — noting that under its limits, the largest legal training run could still complete in about 12 days. It cites the Non-Proliferation Treaty and the Chemical Weapons Convention as precedents.

None of that is vague. It is the most concrete superintelligence ban proposal the organisation has written, and it is 17 months old.

The Superintelligence Ban Site Says "Extinction" 101 Times and "FLOP" Zero

We ran word-boundary counts over both bodies of text. Not a search for themes — a search for the actual words, with the counts normalised per 10,000 words so the two documents can be compared fairly.

The inversion

The superintelligence ban site is dense in threat language and empty of mechanism. The plan is the reverse.

Word	controlai.org (34,114 words)	per 10k	A Narrow Path (23,947 words)	per 10k
extinction	101	29.6	21	8.8
superintelligen*	269	78.9	73	30.5
licence / license*	2	0.6	71	29.6
treaty / treaties	3	0.9	53	22.1
threshold*	8	2.3	52	21.7
enforce*	7	2.1	40	16.7
verify*	5	1.5	20	8.4
inspect*	0	0.0	14	5.8
FLOP*	0	0.0	12	5.0

The tidiest coincidence in the corpus

“Extinction” runs at 29.6 uses per 10,000 words on the campaign site. “Licence” runs at 29.6 per 10,000 in the plan. The same density, in one organisation’s superintelligence ban writing, for the problem and for the remedy — kept in separate documents.

Mechanism words per 10,000 words, campaign site vs the plan it links to

Licence — the plan 29.6

Licence — the campaign site 0.6

Threshold — the plan 21.7

Threshold — the campaign site 2.3

Verify — the plan 8.4

Verify — the campaign site 1.5

FLOP — the plan 5.0

FLOP — the campaign site 0.0

The Superintelligence Ban Needs a Definition. ControlAI Publishes Five

A superintelligence ban needs a boundary. Ours is a counting exercise, so we counted the boundaries. Across the campaign site, “superintelligence” or “superintelligent AI” is given a working definition five separate times, and the five do not describe the same thing.

The five

Page	Definition given	Test it implies
Homepage	“more powerful than any individual, any company, or any nation”	Comparative power
Policy proposals (intro)	“can fully replace and outmatch humans at any task”	Universal task coverage
Policy proposals (legal)	“capable of overpowering or undermining the country’s national security establishment”	National security outcome
Kill switch amendment	“can autonomously compromise national security, escape human oversight, and upend international stability”	Three conjunctive outcomes
Direct Institutional Plan	“systems more intelligent than all of humanity combined”	Aggregate intelligence

Why the spread matters

These are not restatements. A system that outmatches humans at any task is not necessarily one that can overpower a national security establishment, and neither is obviously “more intelligent than all of humanity combined”. The kill-switch wording is stricter still: it joins three conditions with “and”, so a system meeting two of them falls outside it.

The superintelligence ban site says “define”, “defines” or “definition” four times in 34,114 words. It offers five definitions.

A Superintelligence Ban Defined by Outcome Is a Choice, Not an Oversight

Here the fair reading matters, because the superintelligence ban campaign addresses this directly and its answer is not a dodge.

The stated reasoning

The policy page, last updated 24 August 2026, says plainly: “we recommend defining superintelligence by outcome rather than by technical details.” The reasoning given is that an outcome-anchored definition “keeps the law relevant no matter how the technology used to build AI changes.”

That is a real argument, and a familiar one in regulatory drafting. Technology-specific law ages badly. A statute pinned to a FLOP count in 2026 is a statute that algorithmic efficiency can walk straight through — a point A Narrow Path concedes itself when it notes thresholds “should be lowered as necessary, to compensate for more efficient utilization of compute.”

The cost of the choice

But a superintelligence ban written around an outcome has a timing problem that a threshold does not. “Capable of overpowering the national security establishment” is a property you can be confident about after the fact, and must guess at beforehand. The prohibition has to bite before the capability exists, which means someone has to certify absence of a capability nobody can measure.

A Narrow Path solves this the conventional way — with proxies, because proxies are what you can inspect. The campaign layer drops the proxies and keeps the outcome. That is the gap, and it is a drafting gap rather than a bad-faith one.

The Superintelligence Ban Scoreboard Disagrees With Itself

The superintelligence ban campaign’s central evidence of traction is its lawmaker count. Six pages give that count, and they do not agree.

Six pages, six numbers

Page	Briefed	Supporters	Period stated
Homepage	375+ (US, UK, Canada, Germany)	135+ UK, 35+ Canada	since Nov 2024
Policy proposals	20+ US Members of Congress, 150+ federal offices, 350+ UK/Canada/Germany	150+ UK and Canada	“less than two years”
Work in the US	100+ offices, 18 senators and House members	not stated	“less than a year” / “since mid-2025”
Work in the UK	175+ parliamentarians	135+ UK	since Nov 2024
2025 Impact Report	200+ parliamentarians	110+ UK	“little more than a year”
Direct Institutional Plan	not stated	20+ UK parliamentarians	“less than three months”

The specific conflicts

US legislators briefed directly: 18 on one page, 20+ on another. US offices: 100+ against 150+. UK supporters: 110+ in the impact report, 135+ on three other pages. And the homepage’s 135 UK plus 35 Canadian supporters total 170, where the policy page says “over 150 lawmakers in the UK and Canada.”

The US page contradicts itself inside four sentences, saying both “in less than a year” and “since mid-2025” — a period that, by September 2026, is fourteen months.

None of these gaps is large enough to change whether the superintelligence ban campaign is working. They matter because this is the campaign’s own measurement of itself, and a movement whose case rests on counting ought to be countable.

The Superintelligence Ban Site Has Four Pages That Load Empty

ControlAI publishes a sitemap listing 60 URLs. Four of them return a page shell with the title “ControlAI” and no body content at all.

Which four

They are /our-work-in-canada, /risks, /projects and /andrea-miotti. Each returns an identical 178,782-byte response with nothing rendered inside it.

The first is the notable one. Canada appears in the homepage’s headline claim — “375 lawmakers in the US, UK, Canada, and Germany” — and in the 35+ Canadian supporters figure. The page a reader would click to check it is blank. The fourth is the biography page of the organisation’s founder and chief executive.

Smaller things that point the same way

The homepage runs its six-sentence explainer of superintelligent AI twice in succession, ending “the extinction of the humanity” the first time and “the extinction of the human race” the second. The US page repeats the sentence beginning “Led by our US Director Connor Leahy” three times, once truncated mid-clause. The kill-switch amendment page carries the wrong title tag — it is served under the heading of the separate UK bill page.

TechCrunch introduces Leahy as ControlAI’s “U.S. Executive Director”. ControlAI’s own US page calls him “US Director”, three times.

None of this is evidence for or against a superintelligence ban. It is evidence about how much of the campaign’s output is a website and how much is a machine-readable case, which is the question a policymaker’s staffer answers first.

No Superintelligence Ban Has Been Introduced Anywhere Yet

The episode summary says what “sounded far-fetched six months ago is now being backed by a wave of new legislation.” For a superintelligence ban, that is checkable, and it is partly true.

What actually exists

The concrete item is a “kill switch” amendment to the UK’s Cyber Security and Resilience Bill, drafted with Alex Sobel MP and presented in the House of Lords by Lord Clement-Jones on 3 September 2026. It would give the Secretary of State last-resort power to order UK data centres shut down during an AI emergency. ControlAI says it is supported by over 10 cross-party MPs.

Note the vehicle. The first statutory definition of superintelligence is being carried by a cybersecurity statute, because there is no AI bill in front of Parliament to carry it. That is a pragmatic route rather than a designed one, and it constrains what the definition can be asked to do.

It also contains a statutory definition of superintelligence, which — as ControlAI itself notes — “could be the first time that superintelligence is defined in law.” That is a real milestone, and the campaign is right to claim it.

What does not exist yet

ControlAI’s own Artificial Superintelligence Security Bill, drafted with a parliamentary lawyer and endorsed by Sir Stephen Fry, has not been introduced. The site says the organisation is “actively working to introduce” it. In the US, the campaign’s blog points to the Artificial Intelligence Risk Evaluation Act and to California’s SB 53 — neither of which is a superintelligence ban.

One tabled amendment and one unintroduced draft bill is meaningful progress for a two-year-old superintelligence ban campaign. It is not a wave. Worth noting: the word “Congress” appears zero times in the 8,397 words of the campaign’s blog index, which is an odd silence for an organisation whose newest hire leads its US operation.

What a Superintelligence Ban Would Mean for Businesses Using AI

Strip the existential framing off the superintelligence ban and a concrete question sits underneath: if these thresholds became law, who would be caught?

The threshold is not exotic — it is already European law

A Narrow Path sets its Training Licence at 10^25 floating-point operations. That is the same figure as Article 51(2) of the EU AI Act, which presumes a general-purpose model has high-impact capabilities when cumulative training compute exceeds 10^25 FLOP. The superintelligence ban’s headline number is not a novel restriction. It is the line the EU already drew, with a licence attached instead of a disclosure obligation.

For almost every organisation reading this, that threshold is irrelevant — it catches frontier trainers, not deployers. Fine-tuning, retrieval and agent orchestration sit nowhere near it.

The clause that would reach further

The one that would touch ordinary buyers is the data-centre rule: a Compute Licence for facilities above 10^17 FLOP/s, with know-your-customer requirements and physical GPU tracking. That is a supplier-side control with a customer-side consequence, in the same family as export controls — it would show up as provider due-diligence obligations, not as a rule you comply with directly.

The honest summary for a technology buyer

No superintelligence ban changes a deployment this quarter. What it changes is the shape of the regulatory tail. The debate has moved from “how do we align these systems” toward “should some of them exist”, and our earlier analysis of the AI alignment problem as a business risk covers the containment side of the same argument.

What to Ask Before You Take a Side on the Superintelligence Ban

The counting above is not an argument that the superintelligence ban is wrong. It is an argument about which parts of its case are checkable.

Three questions that separate the layers

First: which superintelligence ban document is being cited? A claim sourced to A Narrow Path carries thresholds and a verification regime. The same claim sourced to a campaign page carries neither.

Second: which definition is in play? The five on the site imply different tests, and the one entering UK law is the narrowest.

Third: what would falsify it? The plan’s proxies are falsifiable — a compute threshold is either exceeded or not. An outcome definition is not falsifiable in advance, which is precisely what makes it durable and what makes it hard to enforce.

Where this sits against the rest of the week

This is the third such gap we have counted in three days. Anthropic’s alignment lead put the odds of AI killing everyone above 10% while the company’s 28,894 published safety words never say “extinction”. Anthropic’s own essay on recursive self-improvement risk uses the word “risk” once in 4,944 words. The UN human rights chief used “existential risk” once in a 3,820-word speech.

The pattern holds across labs, superintelligence ban campaigners and regulators alike: the strongest claims travel in the least durable formats, and the operational detail sits in documents almost nobody opens.

The Superintelligence Ban Question the Audio Never Had to Answer

TechCrunch asked whether we should let superintelligence happen. A superintelligence ban is the right question to put to that, and 41 minutes is a reasonable length at which to argue about it.

But the answer a policymaker needs is not an argument. It is a number, a unit, and an inspection regime. The superintelligence ban has all three. They are on page 40-odd of a PDF from April 2025 that the campaign’s own pages mention in passing and quote almost never.

So the finding is narrower than the headline and, we think, more useful. The superintelligence ban case is not unserious — it is drafted, lawyered, and now carries a definition into a live parliamentary amendment. It is that the serious version and the popular version are different documents, and only one of them is 34,114 words of extinction.

Whether superintelligence arrives before 2030, as the homepage says it might, is not something we can check. Which of the five superintelligence ban definitions a government would have to enforce is something anyone can check, and should, before deciding the answer.