Former UK Defence Secretary: Ban Superintelligent AI Before It's Too Late

A former UK Defence Secretary who spent decades working on nuclear non-proliferation has issued one of the starkest warnings yet from the political establishment about artificial intelligence: superintelligent AI now poses a threat "analogous in scale to that of nuclear war," and governments should prohibit its development outright before it is too late.

Lord Des Browne, who served as Defence Secretary from 2006 to 2008 and is now retiring from the House of Lords, makes the case in a commentary published by the Royal United Services Institute (RUSI) on August 13, 2026. He argues that world-leading AI scientists and the CEOs of the top AI companies already acknowledge that a fully developed superintelligent system would pose an existential risk to humanity — and that the companies building it are continuing anyway.

The piece is significant not because the underlying warnings are new, but because of who is making them. Browne spent his career around the most destructive technology humans have ever built, and he is saying AI is worse.

The Emerging Threat

Browne's argument begins with the nuclear precedent — and with the point where the analogy breaks down. In the years before the Manhattan Project's Trinity test, some scientists feared a fission detonation could ignite the atmosphere in a runaway reaction. They ran the calculations, concluded the risk was extraordinarily unlikely, and proceeded. The atmosphere did not ignite.

Today's AI scientists are telling a different story. "The scientists warning us about the extinction risk posed by superintelligence are neither outsiders nor doom mongers," Browne writes. The researchers sounding the alarm are saying there is currently no technical means in sight for humans to control AI systems more intelligent than ourselves. If superintelligence were developed tomorrow, he argues, we would have no way to control it.

The core distinction Browne draws is between a weapon and an agent:

"A nuclear weapon has devastating effects, but it is ultimately an instrument rather than an agent. It cannot improve itself, replicate itself, or resist being shut down. Superintelligent AI systems would be capable of all three, constituting threat actors in and of themselves. In developing such systems, effectively we would be ceding control over the most powerful weapon the world has ever seen to the weapon itself."

This is the heart of the case. Eighty years of nuclear deterrence have succeeded through a combination of statecraft, institutional discipline, and no small amount of luck — and that was with weapons that remain firmly under human command. Browne's point is that a technology capable of self-improvement, self-replication, and resisting shutdown cannot be managed by the same playbook.

The Extinction Risk Warning

Browne is not relying on his own authority alone. He points to the public warnings of Nobel laureate Geoffrey Hinton and Yoshua Bengio, one of the most-cited computer scientists in history, both of whom have repeatedly stated that superintelligence would pose an extinction risk to humanity.

The argument also rests on the collective action of the scientific community. Last year, Browne joined Hinton, Bengio, and more than 800 other prominent figures in signing a call to prohibit superintelligence until there is scientific consensus that it can be developed safely — and widespread public consent. The statement, hosted at superintelligence-statement.org, frames prohibition not as a rejection of AI, but as a pause on the one category of AI that nobody yet knows how to control.

Browne is notably critical of industry in this regard. The CEOs of the leading AI companies, he writes, have acknowledged the extinction risk many times, "though nowadays they prefer to cloud the issue in the hope of avoiding regulation."

Why Existing Oversight Isn't Enough

One of the most important sections of the commentary addresses a question that will occur to anyone following AI governance: don't the UK's existing institutions already handle this?

The UK hosted the world's first AI safety summit in 2023, and it established the AI Safety Institute — now the AI Security Institute (AISI) — before any other nation. But Browne argues that this institutional leadership creates a false sense of security.

AISI, he notes, conducts research and evaluates companies' AI models on a voluntary basis. It has no power to stop a determined actor, inside or outside the UK, from developing a superintelligent system. Research and voluntary evaluation are useful tools for improving model safety; they are not a defense against a technology that can outthink its overseers.

The Case for Prohibition

Browne's proposed response is a two-stage prohibition. First, the UK should prohibit the development of superintelligence on British soil. Second, it should work to build a coalition of states agreeing to a global prohibition.

He argues this is achievable because the threat is universal. "No state has an interest in any actor developing a technology that no one can control, that would strip all states of their sovereignty, and that could extinguish humanity," he writes. In a time of geopolitical upheaval, a shared interest in preventing an uncontrollable technology may be the surest foundation for international cooperation available.

The timing of the piece matters as much as its content. It arrives as the UK government frames AI sovereignty as central to securing Britain's future. Browne's argument is that true AI sovereignty is impossible without human control:

"But true AI sovereignty involves ensuring that humanity remains in control of AI. Prohibiting superintelligence is a sure-shot way to ensure this outcome."

What Critics Would Say

The case for prohibition is not without its counterarguments, and the commentary leaves several of them unaddressed.

The most obvious is enforcement. A prohibition on superintelligence development would be extraordinarily difficult to verify: unlike nuclear enrichment facilities, which require detectable physical infrastructure, AI training runs on distributed compute that can be concealed or relocated. Browne acknowledges the need for a coalition of states but does not explain how cheating would be detected.

There is also the question of definition. "Superintelligence" is a contested term, and a treaty-grade definition that cannot be gamed has yet to be written. And critics of the safety-first position argue that an outright ban would cede leadership to states and companies that ignore it — a version of the same security dilemma that drove the nuclear arms race.

Browne's reply to these concerns is essentially precautionary: the consequences of being wrong in one direction are extinction, and in the other, merely delay. Whether that framing persuades governments is the open question.

What Policymakers and Organizations Should Watch

Whatever the merits of the argument, the commentary signals where parts of the UK establishment are heading:

• Ban proposals are becoming mainstream. A former Defence Secretary publicly endorsing prohibition moves the idea from the fringe of AI safety discourse into legitimate policy territory.
• The sovereignty framing is powerful. By tying superintelligence to national security rather than abstract ethics, Browne gives governments a vocabulary for regulation that fits existing security frameworks.
• Verification will be the battleground. If prohibition gains traction, the debate will quickly shift to how development could be detected and attributed — a problem with no proven technical solution today.
• Industry rhetoric is under scrutiny. Browne's pointed remark about CEOs clouding the issue reflects a growing impatience with voluntary commitments as the primary governance mechanism.

The Bigger Lesson

Browne's commentary is, in one sense, a case study in how existential-risk warnings migrate from scientists to policymakers. Hinton and Bengio warned for years; now a former Defence Secretary is carrying the same message into the security establishment, dressed in the language of sovereignty and national interest.

The nuclear analogy cuts both ways. It gives the warning credibility — but it also reveals how hard the problem is. The world did eventually build an arms-control architecture for nuclear weapons, and it has held, barely. But it took decades, it required mutual deterrence between roughly equal powers, and it has never fully solved the problem of cheating.

Browne's own conclusion is simpler and more urgent. "We have been fortunate thus far not to destroy our world with nuclear weapons," he writes. "But we cannot expect to be fortunate with AI that is beyond our control. The opportunity to avoid engineering our own extinction will not be offered to us a second time."

Whether governments agree to prohibit superintelligence — or whether they decide the risks of banning outweigh the risks of building — the debate has now been joined at the highest level of British politics. The window for having it while humans are still the most intelligent systems in the room may not stay open forever.