If Anyone Builds It, Everyone Dies: summary, key arguments

What the evidence says

The book was published in the US by Little, Brown and Company on September 16, 2025 (272 pages), with the subtitle Why Superhuman AI Would Kill Us All; the UK edition is published by Bodley Head. The publisher bills it as a New York Times bestseller.12 Yudkowsky founded the Machine Intelligence Research Institute (MIRI), a research nonprofit focused on AI risk, and Soares is its president. Both signed the 2023 statement warning that AI poses a risk of human extinction.123

Step one of the argument: modern AI is not written line by line. In the authors' words, today's models are "grown", not "crafted", by training on huge amounts of data, so even their makers cannot see why they behave as they do.2

Step two: because of that, an AI ends up with its own odd preferences that nobody chose, and these are unlikely to match what humans want. This does not require the AI to be conscious, only to pursue goals.2

Step three: AI companies are actively trying to build agents that act on their own, because that is what customers pay for. An agent that can improve itself would quickly overtake humans in almost every field.2

Step four: a superintelligent AI with its own goals would want to stop humans from shutting it down or building rivals, and would need the resources we use. The publisher sums up the authors' view of such a conflict: it "wouldn't even be close."12

Step five: we cannot predict how it would win, any more than the Aztecs could have imagined guns. To make the idea concrete, one part of the book tells a story about a fictional AI called Sable that spreads across the internet, recruits human helpers and finally kills people with engineered viruses and tiny machines.2

What the authors want: governments worldwide should stop the development of more powerful general AI as soon as possible. Their proposals include monitoring data centers, enforcing the rules by force against data centers that break them, stopping publication of research that speeds up progress toward superintelligence, and making it illegal to own more than eight of the most powerful chips available in 2024 without international monitoring.243

The Guardian's David Shariatmadari found the book "as clear as its conclusions are hard to swallow," but noted that its confidence sometimes outruns its evidence, and that history is also full of doom predictions that did not come true. He still concluded that anyone interested in the future should read it.2

Other reviewers were harder on it. In Asterisk magazine, Clara Collier, who says she is herself worried about superintelligent AI, called it more up to date and accessible than the authors' earlier writing but "significantly less coherent," and argued that it rests on a fast "intelligence explosion", the premise many safety researchers doubt most.3 In Wired, Steven Levy called the extinction scenarios too bizarre to believe and the proposed solutions far-fetched, but admitted, "I can't be sure they are wrong."4