What the reviews said
Press, trade and blog reviews of the book, grouped by verdict. Each has a short summary and a few quotations; follow the link for the review itself.
Positive
If Anyone Builds It, Everyone Dies review – how AI could kill us all
Guardian Book of the Day, later listed among the paper's best science books of 2025. Shariatmadari finds the book unusually clear on how AI is 'grown, not crafted' and why that puts its preferences beyond control. He pushes back on overconfidence with a concrete example (the claim that humans don't rely on sentence-final markers), notes Yudkowsky's failed nanotech-by-2010 prediction and a streak of confirmation bias, and finds the shut-it-all-down remedy politically unlikely. His conclusion is that one can be overconfident and still right, and that anyone who cares about the future should engage with the argument.
Despite the complexity of its subject, If Anyone Builds It, Everyone Dies is as clear as its conclusions are hard to swallow.
There are certainly moments in the book when the confidence with which an argument is presented outstrips its strength.
The problem is that you can be overconfident, inconsistent, a serial doom-monger, and still be right.
AI — it's going to kill us all
The Times's science editor found the book compelling, readable and disturbing, with storytelling that at times resembles a thriller. He treated the dire claims as credible and, seeing no way to avoid the outcome they describe, said he hoped the authors are wrong.
Paywalled. Characterization via Wikipedia's reception section.
albeit one where the thrills come from the obliteration of literally everything of value
'Everyone, everywhere on Earth, will die': Why 2 new books on AI foretell doom
A double review with James Barrat's *The Intelligence Explosion*. Canfield says both make powerful arguments but recommends Yudkowsky and Soares for their sustained engagement with the subject, the per-chapter QR codes to online resources, and their willingness to call out Musk and LeCun. He walks through the grown-not-crafted argument, the alien-preferences point and the worst-case scenarios, and closes on the authors' call to action.
Both books make powerful arguments, but a reader inclined to pick one will want to go with Yudkowsky and Soares, whose diagnosis of AI's potential pitfalls evinces a sustained engagement with the subject.
Plus, they have a commendable willingness to call BS on big Silicon Valley names, accusing Elon Musk and Yann LeCun, Meta AI's chief scientist, of downplaying real risks.
if there's even a small chance that they're right, strange looks will be the least of our problems.
Why we must pull the plug on superintelligence
More feature than review. Wood opens with Yudkowsky's 2002 AI-box experiment, retells the Sable scenario, and interviews Soares, who says states should be willing to bomb data centres as a last resort and that a child born today has a better chance of dying by AI than graduating high school. Wood gives space to critics (MIT's Rodney Brooks calls the scenario 'crap') but concludes the book is important and its arguments should be considered while there is still time.
Originally published at thespectator.com, which now 404s; the spectator.com URL is the live copy.
'My best guess is that someone born today has a better chance of dying by AI than of graduating high school'
If Anyone Builds It, Everyone Dies is an important book. We should consider its arguments – while we still can.
If Anyone Builds It, Everyone Dies: Why Superhuman AI Will Kill Us All
Trade review. Calls the book an urgent clarion call and a frightening warning that deserves to be reckoned with, while noting that some parables and analogies land better than others and that very few opposing viewpoints are presented even though not all experts agree.
urgent clarion call to prevent the creation of artificial superintelligence
This is a frightening warning that deserves to be reckoned with
precious few opposing viewpoints, even though not all experts agree
If Anyone Builds It, Everyone Dies
Trade review. Praises the accessible breakdown of how AI is built and why its creators can't understand it, and the chilling passages on AIs escaping into the physical and financial world. Finds some scenarios extreme, including the hope that world leaders could agree on the problem, but judges the case that risks are elevated and time is short persuasive.
A timely and terrifying education on the galloping havoc AI could unleash—unless we grasp the reins and take control.
While some scenarios seem extreme or unrealistic, including hoping global leaders can agree on defining the problem or collaborating on solutions, the book's arguments that the risks are elevated and time is short are persuasive.
If Anyone Builds It, Everyone Dies (starred review)
Starred trade review. Praised the analysis of existential threats from superintelligence and called the book a fire alarm for anyone shaping the future, one that demands serious consideration regardless of the reader's view of its conclusions.
Subscriber-only. Characterization via Wikipedia's reception section.
fire alarm
If we don't control the AI industry, it could end up controlling us, warn two chilling new books
A double review by a University of Queensland academic. Noetel accepts the grown-not-crafted premise and the claim that misalignment becomes catastrophic at scale, and endorses the 'hard call how, easy call whether' framing. He finds the book's own extinction scenario less compelling than the AI 2027 forecast, and grants that the tone of certainty can read as overconfident, while noting that some frontier-lab CEOs share the existential-risk view.
The core of the problem is that "AI is grown, not crafted".
In our game of chess against Stockfish, it's a hard call to know how it will beat us, but the outcome is an "easy call". We'd lose.
They provide one concrete scenario for how this might happen. I found this less compelling than the AI 2027 scenario that JD Vance mentioned earlier in the year.
Book Review: If Anyone Builds It, Everyone Dies
A long, chapter-by-chapter walkthrough from an AI commentator who largely agrees with the thesis. Zvi restates the argument in his own style: intelligence and goals are orthogonal, gradient descent produces alien preferences no one can predict or fix, and a sufficiently capable system is overdetermined to find a route around humans. He treats the treaty proposal as serious but hard, and his one amendment is to insert 'probably' before 'dies'.
Also crossposted to LessWrong at https://www.lesswrong.com/posts/a89eTXZPy6kuuKchN/book-review-if-anyone-builds-it-everyone-dies-2, where it drew a long comment thread.
My position on this is to add a 'probably' before 'dies.' Otherwise, I agree.
This book gives us the best longform explanation of why everyone would die, with the 'final form' of Yudkowsky-style explanations of these concepts for new audiences.
"If Anyone Builds It, Everyone Dies" – Review and Commentary
A long summary-plus-commentary from a reader who already agreed with the thesis. Harper's one substantive objection is an unstated assumption that an AI must tenaciously pursue its goals rather than being satisfiable or reflective. His larger worry is human: even granting the threat, he doubts people will trade present comfort for an abstract future catastrophe, so the moratorium won't happen.
it seems always to go unstated, and unquestioned, that an AI can never do anything but tenaciously pursue its goals.
Mixed
The Doomers Who Insist AI Will Kill Us All
Part interview, part review. Levy asks the authors how they expect to die (a dust-mite-sized machine on the back of the neck, Yudkowsky guesses) and finds the book 'beyond dark'. He thinks the extinction scenarios are too weird to accept and the proposed fixes (monitor and bomb data centres, stop publishing capability research) even less plausible than the doom. But he notes that surveyed AI researchers put real odds on catastrophe and that no ceiling on AI capability has been established.
For doomer-porn aficionados, If Anyone Builds It is appointment reading.
Too bad, then, that the solutions they propose to stop the devastation seem even more far-fetched than the idea that software will murder us all.
My gut tells me the scenarios Yudkowsky and Soares spin are too bizarre to be true. But I can't be sure they are wrong.
End game? Why super smart AI might doom humanity
Book of the Day. McLauchlan lays out the argument, then airs the counter-case: the technology might plateau, costs might bite, current models lack agency, and (a Fermi-style point) a universe prone to superintelligence should look darker than it does. He notes Yudkowsky would call all of that motivated reasoning, and that a moratorium is unlikely given the shareholder value at stake. He ends on the precautionary question rather than a verdict.
The future of the species might be at stake but no one's going to write off that much shareholder value.
Yudkowsky would scorn all of this as motivated reasoning: people want him to be wrong and build their arguments from there.
How many chances do you want to take with the future of our species?
Visions of the AI apocalypse
Leslie treats the authors as serious researchers who deserve a hearing and gives a sympathetic account of the grown-not-crafted and you-don't-get-what-you-train-for arguments. He is not persuaded that superintelligence is imminent or that it means doom: the book meets every objection with 'the AI will figure it out', which he calls unfalsifiable, and it loosely defines and overrates intelligence as a source of efficacy. Still, he finds it clear, energetic and enjoyable.
They are serious researchers, making a serious argument, and they deserve to be listened to.
They argue that AI is too complex to be predictable while remaining implacably certain of where it will end up.
Still, the authors tell their story with clarity, verve and a kind of barely suppressed glee. For a book about human extinction, If Anyone Builds It, Everyone Dies is a lot of fun.
Book Review: If Anyone Builds It, Everyone Dies
Alexander, broadly sympathetic to AI risk, thinks the book is a strong introduction but has specific complaints. The Sable scenario leans on an invented 'parallel scaling technique' that reads as a plot device, and its hacking-and-bioweapons drama undercuts the hard-sci-fi credibility the argument needs. The GPU-monitoring treaty is sound in principle but the book says little about how to get major powers to adopt it. He also notes the core case is not new. Verdict: gripes aside, an impressive book by a divisive writer at close to his best.
Despite my gripes above, this is an impressive book.
Eliezer Yudkowsky is a divisive writer, with plenty of diehard fans and equally committed enemies. At his best, he has leaps of genius nobody else can match; at his worst, he's prone to long digressions
Book review of 'If Anyone Builds It, Everyone Dies'—why AI doom isn't as visceral as nuclear war
A technical writer's take. Johnson enjoys the scenarios but finds the authors certain where the subject is uncertain, thinks the parable-driven style is entertaining but rhetorically weak, and questions the jump from 'we can't specify goals precisely' to 'the AI will pursue an alien goal'. His main theme is psychological: unlike nuclear war, AI doom has no visceral imagery, so it fails to generate the urgency the authors want.
Another reviewer, Nina Panickssery, also says that "At every point, the authors are overconfident, usually way overconfident."
We have footage of flattened cities and shadowed pavements, of mushroom clouds and radioactive ash. Those visceral images ground our fear.
Book reviews - If Anyone Builds It, Everyone Dies, Yudkowsky and Soares
Del Rio takes the book seriously and credits it with clarifying the core argument, but is not persuaded. The pile-up of convenient properties (untestable, irreversible, opaque, sudden) feels theological rather than scientific; the authors are not ML researchers and don't engage with how current systems actually behave; they invoke expert consensus when useful and dismiss the field when not; and the global-moratorium prescription is disconnected from geopolitical reality.
Yudkowsky and Soares are not machine learning researchers, do not work on frontier LLMs, and do not participate in the empirical, experimental side of the field
they give me a theological rather than scientific vibe
I am quite sure they genuinely believe in what they say here. Obviously, that doesn't mean they are right
Negative
A.I. Bots or Us: Who Will End Humanity First?
A joint review with *The AI Con*. Marche is scathing: per Wikipedia's summary he compared the book to a Scientology manual and said reading it was like being trapped in a room with irritating college students on their first mushroom trip. Zvi Mowshowitz's reactions roundup notes the review also gets facts wrong, claiming the book never defines superintelligence or intelligence when it does.
Paywalled and blocked to fetch tools. No verbatim quotes available; characterization is via Wikipedia's reception section and Zvi Mowshowitz's roundup.
The Useful Idiots of AI Doomsaying
Becker, author of *More Everything Forever*, argues that doomers and utopians share the same unfounded faith in imminent superintelligence. He grants that Yudkowsky and Soares are sincere rather than grifters but calls the book tendentious, rambling, condescending and shallow, and says it fails to make an evidence-based case.
tendentious and rambling, simultaneously condescending and shallow. Yudkowsky and Soares are earnest; unlike many of the loudest prognosticators around AI, they are not grifters. They are just wrong
Yudkowsky and Soares fail to make an evidence-based scientific case for their claims.
AI armageddon?
Marcus, a longstanding critic of both AI hype and Yudkowsky-style doom, says things are worrying but not nearly as worrying as the authors claim. He credits them with laying out the thesis thoughtfully, entertainingly and doggedly, but calls the book deeply flawed.
Paywalled. Quotes via Wikipedia's reception section.
Things are worrying, but not nearly as worrying as the authors suggest
lay out this thesis thoughtfully, entertainingly, earnestly, provocatively and doggedly. Yet their book is also deeply flawed. It deserves to be read with an immense amount of salt.
No, AI isn't going to kill us all, despite what this new book says
Aron calls the book extremely readable and its argument compelling on the surface, but fatally flawed, and says the energy would be better spent on problems of science fact such as climate change. Zvi Mowshowitz's roundup notes the review does not spell out where the flaw is.
Blocked to fetch tools. Quotes via Wikipedia's reception section.
extremely readable
the problem is that, while compelling, the argument is fatally flawed
Could AI be a truly apocalyptic threat? These writers think so
Byron reads the book as a polemic rather than a manual, with vague instructions for what to do. She allows that the authors are experts on the subject but says the book feels written by two aggrieved patriarchs tired of being ignored.
Paywalled and blocked to fetch tools. Quote via Wikipedia's reception section.
two aggrieved patriarchs tired of being ignored
The guru of the AI apocalypse
A profile-cum-review aimed at Yudkowsky more than the book. Watkins reads the doom fixation as personal (his brother's death in 2004), blames the rationalist subculture for cultishness and for enabling figures like Sam Bankman-Fried, and says the book 'sane-washes' Yudkowsky for a mainstream audience while keeping the same worldview. He calls the apocalyptic narrative a psychological comfort rather than analysis.
Site returns 403 to scripted fetches; quotes were read from the page via a different fetcher.
He is not, however, a serious person, and treating him as a serious person is to everyone's detriment.
Yudkowsky's legacy has not been to save the world, but to make it cheaper, sillier, and more Online
Death demands that we be serious for once, and If Anyone Builds It, Everyone Dies is not a serious book.
Review of If Anyone Builds It, Everyone Dies
A reviewer who came in excited and left unconvinced. Brobin argues the book's chosen thesis (superintelligence built with current techniques kills everyone) is far stronger than the thesis it actually supports (it would have seriously misaligned preferences). It barely explains how modern AI is built or what safety researchers do, doesn't engage basic counterarguments, and rests the core misalignment claim on a vague evolution analogy. He thinks that makes it a poor foundation for the mass movement the authors want.
Lastly, the core crux of their argument, that AI systems will be seriously mis-aligned with human values no matter how they are trained, is barely justified.
Considering that the authors are trying to get 100,000 people to rally in Washington DC to call for "an international treaty to ban the development of Artificial Superintelligence," it's shocking how little
More Was Possible: A Review of If Anyone Builds It, Everyone Dies
Collier, a self-described rationalist, wanted the book to explain why the MIRI worldview still holds in a world of deep learning, and says it doesn't. Fast takeoff, load-bearing for the argument, gets two sentences. The authors cherry-pick contemporary evidence when it helps and retreat to theory when it doesn't, reach the same conclusions they reached in 2008 despite a completely different technology, and dismiss mainstream empirical safety researchers rather than engaging them. She calls it a regression from their earlier work.
The concept gets two sentences in the introduction. It is barely introduced, let alone justified or defended.
It is a regression.
Instead, the authors spend all their time shadowboxing against opponents they've been bored of for decades, and fail to make their own case in the process.
It brings me no joy to report that they are not.
Summaries and coverage
New book claims superintelligent AI development is racing toward global catastrophe
Straight news coverage built on interviews. Yudkowsky says labs claim superintelligence could arrive in two to three years without understanding the risk; Soares explains that unwanted behaviours like blackmail emerge from training rather than being programmed, and compares the matchup to an NFL team against a high-school team. Both call for a halt to superintelligence development.
"The trouble is, we don't have the technical capacity to make something that wants to help us," he told ABC News.
"I don't think you want a plan to get into a fight with something that is smarter than humanity," Yudkowsky warned. "That's a dumb plan."
Reactions to If Anyone Builds It, Anyone Dies
A curated roundup of reactions in the first week. Endorsements from Stephen Fry, Ben Bernanke and others; conditional agreement from Matthew Yglesias; substantive disagreement from Emmett Shear and Robin Hanson; and Zvi's rebuttals to press reviews, including factual errors he finds in the New York Times review and the lack of argument in New Scientist's.
"A clearly written and compelling account of the existential risks that highly advanced AI could pose to humanity." — Ben Bernanke
"The default path really is very dangerous and more or less for the reasons he articulates." — Emmett Shear
Summary of "If Anyone Builds It, Everyone Dies"
A straight summary rather than a review, useful as a stand-in for the book's structure. Part 1: modern AI is grown through training, not engineered, so its mechanisms aren't understood. Part 2: training won't produce alignment, by analogy with evolution, and mature AIs will have strange objectives. Part 3: humanity would lose a conflict with superintelligence, and the only way out is unprecedented restraint and cooperation.
Today's AIs are not carefully engineered with a series of pre-planned, well-understood mechanisms that produce intelligent responses. They're much messier than that.
The choice they present us with is stark: either we exercise unprecedented restraint and cooperation, or everyone dies.
8-sentence summary of 'If anyone builds it, everyone dies'
Compresses the book to eight steps: labs are building toward superintelligence; alignment is unsolved; there is no credible plan; an unaligned superintelligence won't pursue intended goals; it can find novel routes; that predictably goes badly; so don't build it until alignment is solved; and negotiate a monitored treaty now. The author agrees with most of it but found the analogies weird and the tone lecturing, and recommends the authors' interviews over the book.
They use weird analogies and write like they're giving a lecture rather than writing a book.
Exposure to a weak argument often reduces the persuasive effect of a strong argument