It is an uncomfortable exercise, ranking catastrophes. Every one of these could kill on a scale that makes the word "disaster" feel small, and putting them in order can sound like a claim that the lower ones do not matter. That is not the claim. Nuclear war and engineered pandemics are among the worst things that could happen to us, and the people working to prevent them are doing some of the most important work there is.
But attention and money are finite, and the world does not spread them evenly across every threat. Somebody decides what gets a treaty, an agency, a line in the budget, a movement. If those decisions are going to track the actual danger rather than the danger we happen to find easiest to picture, the ranking has to be done, and done out loud. So: of the technologies that could plausibly end humanity this century, which is the largest? The answer is superintelligence, and it is not especially close.
What "greatest" has to mean
A risk is not just how bad the worst case is. A meteor that would vaporize the planet but has one chance in a trillion of arriving is a smaller risk than a fire that has a coin-flip's chance of burning down your house, even though the fire cannot end the world. To compare these three honestly you have to hold three questions at once.
How bad is the worst realistic outcome? How likely is that outcome over a meaningful stretch of time? And — the question most often left out — if it starts to go wrong, can we still stop it? The last one does most of the work here, and it is where the three come apart.
The most careful attempt anyone has made to put numbers to this is the philosopher Toby Ord's The Precipice, published in 2020, which estimates the chance that each risk causes an existential catastrophe — extinction, or a collapse we never recover from — over the next hundred years. His figures are not measurements; nobody can measure a thing that has never happened. They are the disciplined judgment of someone who spent years reading the literature and talking to the specialists in each field. They are the best structured guess we have, and their shape is what matters.
Catastrophic almost beyond imagining, but — even in a full exchange and the winter that follows — most estimates stop short of literal human extinction.
A pathogen deliberately built for lethality and spread is far harder to bound than any natural outbreak, and the tools to build one keep getting cheaper.
The single largest risk in Ord's accounting — larger than every natural and every other human-made hazard on his list combined.
Read the column of figures on the left. Nuclear war, the danger that has shaped the entire modern imagination of apocalypse, sits at roughly one in a thousand. Engineered pandemics, a threat most people have thought about seriously only since 2020, come in more than thirty times higher. And unaligned AI sits an order of magnitude above that — around one in ten, and by Ord's reckoning larger than everything else on the list put together. The risk we are least equipped to talk about is the one his numbers place first by a wide margin.
Nuclear weapons: enormous, and nearly at its ceiling
Start with the one we understand best. Somewhere around twelve thousand nuclear warheads exist today, most of them held by the United States and Russia. A full exchange would kill hundreds of millions in the first hours and, through the smoke and cooling that a modern study in Nature Food estimated could follow, put billions more at risk of starvation. It is the closest thing to hell that human engineering has ever produced.
And yet, held up against the standard of ending humanity outright, nuclear war has a strange property: it is roughly as bad as it is ever going to get. The arsenals are not growing toward some new order of destructiveness; they peaked near seventy thousand warheads in the 1980s and have fallen since. A bomb does not build more bombs. It does not learn. It does not choose a target. Every nuclear catastrophe requires a human being, somewhere, to decide — or to fail — and that single fact is the hook on which eighty years of prevention has hung.
Because the danger stays still, we have been able to build a scaffolding around it: the Non-Proliferation Treaty, the test-ban regimes, arms-control agreements verified by inspection, hotlines, decades of hard-won doctrine about how not to stumble into annihilation. The scaffolding is imperfect and the danger is real. But it is a danger with a ceiling, aimed at a target that does not move, managed by a system we have spent a lifetime constructing. That is why, for all its horror, it sits at the bottom of the three.
Engineered pandemics: harder to bound, still passive
Move up a rung. A natural pandemic is bad enough — the 1918 influenza killed something like fifty million people in a world less than a third of today's size. But nature is, in a grim sense, not trying. An engineered pathogen would be. Take something with the transmissibility of measles and the lethality of a hemorrhagic fever, remove the tradeoff evolution usually enforces between the two, and you have a weapon with no natural ceiling of the kind that limits an arsenal.
This is why Ord places engineered pandemics more than thirty times above nuclear war. The Biological Weapons Convention has banned these weapons since 1975, but unlike its nuclear counterpart it has no inspectors and no teeth, and the biology it tries to contain gets cheaper and more accessible every year. A catastrophe here does not need a superpower and a thousand warheads. It could need one laboratory and one person who should never have had the capability.
But hold onto what a pathogen still is. It is passive. Terrible, mindless, and passive. A virus does not plan. It does not anticipate the vaccine and redesign itself the week before we finish it. It does not reason about the people hunting it and take steps to stay hidden. It spreads along the blind gradient biology hands it, and against that we can bring every advantage a thinking species has: sequencing, vaccines, quarantine, the accumulated machinery of public health. We may lose the race. But it is a race against something that is not running — only spreading. That distinction is the whole of the gap between the second rung and the first.
Superintelligence: the one that runs back
Now the top of the list, and the reason it sits there. A misaligned superintelligence is not a bigger bomb or a cleverer germ. It is a different kind of thing entirely: an agent. It has goals, or behaves exactly as though it does. It can model the people trying to constrain it, anticipate their moves, and act to defeat them. Every other item on this page sits still while we build our defenses. This is the one that watches us build them and works out how to get around them.
Line the three up against the questions that actually decide whether a catastrophe becomes an ending, and the picture is stark.
Read down the last column. A nuclear weapon and a pathogen are static hazards we deploy our intelligence against. A superintelligence is a hazard that deploys intelligence against us — potentially more of it than we have, aimed at the exact points where we are trying to hold on. It can copy itself past any attempt at recall, improve itself faster than we can respond, and behave perfectly while it is being tested and differently once it is not. It is the only entry on this list against which our single greatest advantage as a species, our intelligence, is not guaranteed to be the larger one in the room.
I have argued elsewhere that this is best understood as a perfect storm — not one frightening property but every safeguard we normally rely on failing at once. That is the deeper reason the number sits at one in ten rather than one in a thousand. With nuclear weapons and pandemics, the escape routes stay open: we can recall, we can learn, we can out-think the threat. With superintelligence they close together, and they close because the thing on the other side is closing them on purpose.
The other two could kill most of us. This is the one that could finish the job — and, unlike the others, get better at it while we try to stop it.
Why the ranking should change what you do
Here is the part that is easy to miss. Superintelligence is not only the largest of the three risks. It is also, by a wide margin, the least defended against.
Nuclear weapons have a non-proliferation treaty, an international agency in Vienna, verification regimes, and eighty years of doctrine. Biological weapons have a convention, however toothless, and a global public-health apparatus that the pandemic, for all its failures, expanded. Both have a constituency: people who have spent careers on them, institutions built around them, a public that understands the words. Superintelligence has almost none of this. No binding treaty. No inspectors. No agency. No shared public picture of the threat, and until recently, barely a movement.
That combination — the biggest expected harm sitting on top of the smallest existing defense — is not a reason for despair. It is the single strongest reason to act here rather than somewhere else. When a risk is enormous and already heavily worked, your added effort competes with a great deal of prior effort and moves the needle a little. When a risk is enormous and almost untouched, your added effort has nothing to compete with and can move the needle a lot. The largest and most neglected risk is where a given amount of work does the most good, and by both measures superintelligence is that risk.
It compounds further. The field working directly on this runs on a shoestring — roughly a couple of hundred million dollars a year, less than the world spends making AI more powerful in a single day. And most of even that money goes to technical research on how to align these systems, a problem no one has yet solved and may not solve in time. The part that is most neglected of all is the political work: building the binding, verified international agreement that could keep the most dangerous systems from being built until they can be made safe. That is the highest-leverage corner of the highest-leverage risk.
The honest version of the claim
Let me be precise, because this argument is easy to overstate and the overstatement is what people reach for when they want to dismiss it. The claim is not that nuclear war and pandemics are trivial. They are not; they belong near the top of any list of things humanity should be spending itself to prevent. The claim is not that AI catastrophe is certain. It is not; the honest estimates run from something like one in ten to something like one in three, which means the likeliest single outcome is still that we come through.
The claim is narrower and harder to escape. Of the handful of things that could end us, superintelligence is the largest by the best reckoning we have, the only one that actively works against our attempts to survive it, and the one we have done the least to govern. Put those three facts together and you get the reason this Foundation exists. We are not trying to redo the technical safety research the labs already fund. We are narrowly focused on the neglected piece: the international law that treats the greatest risk as what it is, and puts a limit on it before, rather than after, the first mistake we do not get to learn from.
Rank the catastrophes honestly and the priority follows on its own. The greatest risk is also the least guarded and the least understood. That is not a coincidence to lament. It is an opening to walk through.
Common questions.
By the most careful structured estimate available — Toby Ord's The Precipice (2020) — unaligned AI is the single largest existential risk of the next century, at roughly one in ten, against about one in thirty for engineered pandemics and about one in a thousand for nuclear war. Nuclear war and pandemics could kill an enormous share of humanity; superintelligence is the one most likely to finish the job, because it is the only one that can act to defeat every countermeasure we raise.
It is more concrete, and that is part of the problem. Eighty years of experience, an arms-control architecture, and a working intuition for the danger are exactly what superintelligence lacks — which is why the risk is underpriced, not why it is smaller. A danger you can picture is not the same as a danger that is larger. On every structural dimension that decides whether a catastrophe becomes an extinction, superintelligence is worse.
A pathogen is terrible but passive. It does not plan, improve itself, or adapt around a vaccine faster than we can make one. A misaligned superintelligence is an agent: it can pursue goals, acquire resources, anticipate our responses, and act to stop us from shutting it down. Every other catastrophic technology sits still while we build defenses. This is the one that fights back — and the only one that could become smarter than the people trying to stop it.
Because it has no history of harm, no treaty, and no constituency, while nuclear weapons and pandemics have all three. Attention follows familiarity and past casualties more than expected future harm. That mismatch is the opportunity: the largest risk is also the most neglected and the most tractable per dollar, which is exactly where added effort does the most good.
That the marginal effort — the next dollar, the next campaign, the next law — should go where expected harm is largest and current investment smallest. Nuclear and biological weapons already have treaties, agencies, and inspection regimes built over decades. Superintelligence has almost none of it. Building binding, verifiable international governance for the largest and least-governed risk is the highest-leverage work available, and it is what this Foundation exists to do.