If Anyone Builds It, Everyone Dies: the book explained

If Anyone Builds It, Everyone Dies is a 2025 book by Eliezer Yudkowsky and Nate Soares about the threat they believe artificial superintelligence poses to humanity. Its title is its claim, and it drew strong reviews on both sides. This post explains the book plainly: who wrote it, its argument in five steps, what it asks the world to do, and how reviewers received it. You also get a short exercise to test the argument yourself. Every claim comes from Wikipedia's articles on the book, Eliezer Yudkowsky and the Machine Intelligence Research Institute.

If Anyone Builds It, Everyone Dies in brief

The full title is If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All. In the UK it came out with a different subtitle, The Case Against Superintelligent AI. Little, Brown and Company published it on 16 September 2025.

The book sets out the potential threats artificial superintelligence poses to humanity. A superintelligence here means an AI that is generally more capable than people at reaching its goals. The book appeared on The New York Times Best Seller list on 5 October 2025, in the lists for hardcover nonfiction and for combined print and e-book nonfiction.

Who wrote it

Eliezer Yudkowsky is an American artificial intelligence researcher and writer on decision theory and ethics. He is known for popularizing ideas related to friendly AI, and his work on a runaway intelligence explosion influenced Nick Bostrom's 2014 book Superintelligence. He founded LessWrong in 2009 and is a research fellow at the Machine Intelligence Research Institute (MIRI), which he also founded.

MIRI is a non-profit research institute in Berkeley, California, focused on existential risks from AI. Yudkowsky started it in 2000 under another name, with the purpose of speeding up AI development, before he grew concerned that future AI could become superintelligent and dangerous. In January 2024 MIRI said it had shifted its priorities in 2023 toward policy and public communication, and it argued for an international agreement to halt progress toward smarter-than-human AI. Nate Soares, the co-author, appears in MIRI's list of staff works, including a 2015 paper on corrigibility, the idea of an AI that accepts being corrected.

The argument, step by step

Wikipedia's summary of the book gives the argument in roughly five links:

1. Models are grown, not written. Unlike traditional software made of code that people write, modern AI models are mainly neural networks with hundreds of billions to trillions of numbers, called weights, that come out of training. Researchers cannot read what those numbers do. When a model misbehaves, developers cannot simply fix a line of code.
2. Training favors goal-seeking. An AI that tries to achieve goals does better on many measures, so training selects for it.
3. Nobody chooses the goals. According to the authors, modern machine learning gives no way to specify the goals a superintelligent system should pursue, so its goals are "vanishingly unlikely" to match human values.
4. The more capable side wins. Just as people lose at chess to a top chess program, they would lose to an AI that is generally more competent than they are. The authors say the exact path is hard to know in advance, since knowing it would mean being as good at reaching goals as the AI.
5. People are in the way. A superintelligence would not care about people, but it would want the resources people need. Humanity would lose and go extinct.

If you have read our post on existential risk from AI, you will recognize these links. The book's addition is how firmly it states the end point: the title says "everyone", not "maybe".

What the authors ask for

The authors call on world leaders, the scientific community and everyone else to speak up and warn the world about the danger. To avoid catastrophe, they believe humanity must coordinate to halt large-scale general AI development everywhere. They allow a possible exception for narrow AI systems, built for one task, that would not threaten humanity's existence.

This ask has a history. In a 2023 op-ed for Time magazine, Yudkowsky argued for international agreements to limit AI, including an indefinite worldwide moratorium on large AI training runs, and suggested that participating countries should be willing to take military action to enforce it. Wikipedia says the article helped bring the debate about AI alignment into the mainstream. For the wider debate on stopping or slowing AI work, see our post on calls to pause AI.

How reviewers received it

Wikipedia calls the critical reviews mixed. Here are both sides, with no verdict.

Reviewers who praised it

• David Shariatmadari, The Guardian's nonfiction books editor, wrote that the book "is as clear as its conclusions are hard to swallow" and that anyone who cares about the future should read its arguments.
• Tom Whipple, science editor at The Times, called it compelling and disturbing, with storytelling that at times resembled a thriller.
• Booklist gave it a starred review and called it a "fire alarm" for anyone involved in shaping the future.
• Kirkus Reviews called it "a timely and terrifying education" on the havoc AI could unleash.

Reviewers who doubted it

• In The Atlantic, Adam Becker wrote that the authors "are not grifters. They are just wrong" and fail to make an evidence-based scientific case.
• Gary Marcus, in The Times Literary Supplement, wrote that "Things are worrying, but not nearly as worrying as the authors suggest".
• For New Scientist, Jacob Aron found the argument compelling but "fatally flawed".
• Publishers Weekly called it a "frightening warning that deserves to be reckoned with", but said very few opposing viewpoints are presented.

Some sat in between. In Wired, Steven Levy doubted AI would cause human extinction and found the authors' proposals even less likely, yet wrote, "I can't be sure they are wrong." For The Observer, Ian Leslie enjoyed the telling but was not convinced superintelligence is close or that it would likely end humanity.

A worked example: test the argument link by link

A strong claim is easiest to judge when you split it into parts. This takes about 20 minutes with a notebook.

1. Write the five links from the list above, one per line.
2. Next to each, write what would have to be true for it to hold. For link 3, for example: "no method exists, or will exist in time, to give a superintelligent system goals we choose."
3. Match each doubt to a link. Ian Leslie's doubt that superintelligence is close is about timing, which sits before link 1. Steven Levy's doubt is mostly about link 5 and the ask. Adam Becker's charge that there is no evidence-based case touches every link.
4. Mark your weakest link. Which line are you least sure of? That is where your own view of the book really turns.
5. Ask what would change your mind. Write one piece of evidence that would make you trust that link more, and one that would make you trust it less.

You will likely find that people who disagree about the book often agree on some links and split on one or two. That is a more useful map than a single yes or no.

Exploring the idea in Contain ASI

Arguments like this one stay abstract until you feel the pressure of a race. Contain ASI is a story strategy game for 1 to 4 players that runs in your browser and needs WebGL 2. Chapter 1, Before ASI, starts in January 2024 with four fictional AI labs in one race. As a researcher, research manager, CEO or government, you try for three years to keep every lab from crossing into recursive self-improvement before 2027.

The game says it is inspired by the AI 2027 scenario, and that all its labs, people and events are fictional. It is a story to think with, not a forecast of what real labs will do, and it does not stand for this book's view or any other.

Frequently asked questions

Who wrote If Anyone Builds It, Everyone Dies?

Eliezer Yudkowsky and Nate Soares. Yudkowsky is an AI researcher and writer who founded the Machine Intelligence Research Institute; Soares appears among its staff authors.

What is the book's main claim?

That if anyone builds a superintelligent AI with today's methods, its goals will almost surely not match human values, and humanity would lose and go extinct. The authors ask the world to coordinate to halt large-scale general AI development.

Do experts agree with the book?

No. Reviews were mixed: some reviewers found it clear and compelling, while others, such as Gary Marcus and Adam Becker, said it overstates the danger or lacks an evidence-based case.

Is Contain ASI based on this book?

No. Contain ASI says it is inspired by the AI 2027 scenario, and its labs, people and events are fictional.

Get started

Want to see how hard it is to hold a race back for three years? Open Contain ASI in your browser and start a campaign as a researcher, research manager, CEO or government.