Few books state their conclusion in the title. If Anyone Builds It, Everyone Dies, published in 2025 by Eliezer Yudkowsky and Nate Soares of the Machine Intelligence Research Institute, does exactly that. The claim is that if any group builds superintelligent AI using anything like today's methods, the result is human extinction, not as a worst case but as the default. It is the most uncompromising position in the field, presented for a general reader, and it has pushed the argument into mainstream view.
Whatever you make of the conclusion, the book is worth understanding, because it lays out the pessimistic case more plainly than anything else available.
The argument in three moves
The book is structured in three parts, and the argument runs roughly as follows.
First, the case for danger. We are on track to build AI that is far more capable than humans. We do not know how to give such a system goals that are genuinely aligned with human survival and flourishing, and by default a superintelligent system pursuing almost any goal we can currently specify would treat humanity as an obstacle or a resource. The authors lean on the same mechanisms we cover elsewhere, that capability does not imply benevolence, that a capable optimiser resists correction, and that we cannot verify what a system actually wants. Their distinctive claim is how confident they are that these problems are not close to solved and that failure is lethal rather than merely bad.
Second, a concrete story. Rather than leave the danger abstract, the book narrates a specific scenario of how a superhuman AI might come to power and end human control. The authors are explicit that any single story is unlikely in its particulars; the point is to make the abstract mechanism vivid.
Third, what to do. Their prescription is as blunt as the diagnosis: an international halt on the development of superintelligent AI, enforced seriously, including through controls on the computing hardware required to train frontier systems, maintained until the alignment problem is genuinely solved rather than merely hoped over.
The book's core claim is not that AI might go wrong. It is that, on the current path, going wrong is what should be expected, and that going wrong means everyone.
The strongest objections
Serious readers push back, and the pushback is worth stating without strawmanning.
The most common objection is to the certainty. Critics accept that AI risk is real and serious while rejecting the near-inevitability of extinction, arguing that the authors treat contested premises, about how fast capabilities will grow, how hard alignment is, and how a takeover would unfold, with more confidence than the evidence supports. Others argue that alignment progress is more promising than the book allows, or that the very bleakness is counterproductive, pushing people toward fatalism rather than action.
These are legitimate, and the honest response is that the book's confidence is its most contestable feature. You can find the certainty overstated and the argument still deeply unsettling, because the objections mostly attack how sure we should be, not whether the risk is real.
Where the Foundation lands
We share the book's central conviction: that we do not know how to align a superintelligence, that building one on the current trajectory is an unacceptable gamble, and that the responsible response is to stop until the problem is solved. Our own writing on why alignment may not be solvable and why unaligned superintelligence does not deliver the good future reaches similar ground.
Where we differ is in emphasis and tone. The book is close to despairing about whether the alignment problem can be solved in time. We put our weight on the achievability of the solution the book itself endorses, a binding international halt, and treat it not as a forlorn hope but as a concrete political and legal project with real precedents. The Statement on Superintelligence, signed by many of the field's own founders, shows that the call for a halt has moved into the mainstream. You do not need to accept every claim in If Anyone Builds It, Everyone Dies to conclude that its prescription is the sane one. The expected-value case, laid out in our piece on P(doom), gets you there even at far lower confidence than the authors hold. Turning that prescription into reality is the whole of our plan.
Common questions.
It is a 2025 book by Eliezer Yudkowsky and Nate Soares of the Machine Intelligence Research Institute arguing that if any group builds superintelligent AI using anything like today's methods, the result would be human extinction by default rather than as a remote worst case. The book makes the case for the danger, tells a concrete story of how a superhuman AI might seize control, and argues for an internationally enforced halt on developing superintelligence until the alignment problem is genuinely solved.
It was written by Eliezer Yudkowsky and Nate Soares, both of the Machine Intelligence Research Institute. Yudkowsky is one of the earliest and most prominent voices warning about existential risk from artificial intelligence, and the book presents the pessimistic case he and Soares hold for a general audience rather than a technical one.
Its prescription matches the bluntness of its diagnosis: an international halt on the development of superintelligent AI, enforced seriously, including through controls on the specialised computing hardware needed to train frontier systems, and maintained until the alignment problem is actually solved rather than merely hoped over. The authors argue that partial or voluntary measures are inadequate given the stakes.
The most common criticism targets its certainty. Many readers accept that AI risk is real and serious but reject the near-inevitability of extinction, arguing the authors treat contested premises about capability growth, the difficulty of alignment, and how a takeover would unfold with more confidence than the evidence supports. Others contend that alignment progress is more promising than the book allows, or that its bleak tone encourages fatalism rather than action. These objections mostly dispute how certain we should be, not whether the risk is genuine.