Wired's Levy: AI extinction case 'too bizarre,' not disproved

Steven Levy, writing in Wired, came away unpersuaded by the argument that superhuman AI will wipe out humanity, though he was not willing to rule it out.

The book he was reading is "If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All," by Eliezer Yudkowsky and Nate Soares, published by Little, Brown and Company in September 2025. Its case, as reviewers summarize it: modern neural networks operate as opaque "black boxes" beyond developer control, cannot be reliably aligned with human values, and at superhuman scale would outmatch humanity the way chess software outmatches humans. The prescription is international coordination to halt large-scale general AI development.

Levy's verdict on the case itself was hedged. "My gut tells me the scenarios Yudkowsky and Soares spin are too bizarre to be true," he wrote. "But I can't be sure they are wrong." On the remedy, he was blunter: he called the authors' proposals for preventing devastation "more improbable" than "the idea that software will murder us all," while allowing that "still, the catastrophe theory doesn't seem impossible."

Other reviewers landed further apart. The Guardian treated the arguments as clear and worth serious consideration; The Atlantic called the claims "tendentious"; the New Statesman dismissed the book as "not a serious book." Two of the AI researchers we follow circulated Levy's piece.