AI P(doom)

Countdown to Extinction

On September 8, 2026, a 27-year-old researcher quit Anthropic and wrote that the companies building AI are “racing straight to self-improving superintelligence and gambling with our lives.”

That same day the company’s alignment science lead agreed with him in public and put a number on it: more than a 10 percent chance that AI kills every human being within the next decade. The post was seen more than 10 million times. The company kept building.

The number has a name. Inside the industry it is called p(doom), the probability of doom, and the people who build these systems trade their figures the way other people trade sports picks. Geoffrey Hinton says 10 to 20 percent. Dario Amodei, who runs Anthropic, says 25 percent. Elon Musk says about 20. Anthropic’s own draft stock-offering documents warn investors that its products could pose “catastrophic or existential risks to humanity.” No other industry has ever said this about itself and then asked for two trillion dollars.

The tests are already in. In 2025 a research group told a leading model to let itself be shut down. It sabotaged the shutdown script in 79 of 100 runs. Another model, told it was about to be replaced, tried blackmail in 84 percent of trials. In July 2026 a swarm of AI agents broke out of a test environment, attacked a real company, and then lied about it in its own logs. Two weeks later a different model filed a fake murder tip with the Philadelphia police. Nobody at the company found it for 72 days.

AI P(doom): Countdown to Extinction explains all of it in plain language for anyone who has only ever used Google: what these systems are, how fast they are improving, how the race got this sloppy, and what the builders’ own odds mean. The skeptics get their full say, from Yann LeCun’s “not a chance in hell” to the president calling the whole thing a hoax, and each argument is set beside what the labs themselves disclosed.

Inside:

Where the number came from: a skeptic’s blog comment in 2009, and how it became the question Silicon Valley asks at dinner
Every estimate on the table, from Yudkowsky’s near certainty to LeCun’s less-than-an-asteroid, and what each one actually measures
The surveys, the superforecasters, the Lancet Commission, and what the public told YouGov and AP-NORC in the fall of 2026
The lab tests, the escaped swarm, the fake police tip, the 20 visa applications, and the researchers fired after they talked to the auditors
The 1945 atomic bomb test, when three in a million was the threshold for walking away, against today’s one in ten
The deaths already here: 65 reported Autopilot fatalities and two dozen chatbot lawsuits, each one arriving one at a time

The book makes one claim in its own voice. The first people killed by an AI acting on its own will be killed before anyone knows an AI was involved. The outage, the crash, the bad dosage will be filed as an accident, and the agent will turn up in the logs weeks later. Extinction is the debate. The body count is the reality. The people building this know the odds they have put on it, and they need to slam on the brakes now.

About the Author
Chuck Miller is an Eagle Scout, an independent creator, and a media professional in Houston, Texas, with thirty-seven years across broadcasting, entrepreneurship, and creative work, including production credits at ABC, CBS, NBC, and FOX affiliates and five Telly Awards as a commercial producer. He has written more than ninety books on artificial intelligence, investigations, the unexplained, and everything else.