AI CEOs Put Their Own Odds of Catastrophe as High as 25 Percent. The Industry Is Worth $2.37 Trillion Anyway

The term is p(doom), shorthand for the probability that advanced AI causes a catastrophic or extinction-level outcome for humanity. It came out of the rationalist and AI safety communities and moved into mainstream conversation in 2023, after GPT-4's release prompted several prominent researchers to start putting actual numbers on the risk in public.

The numbers vary enormously, and they come from the people who would know best.

Dario Amodei, chief executive of Anthropic, has repeatedly put his own estimate at 10 to 25 percent, telling one interviewer in 2025 there is a 25 percent chance things go, in his words, really, really badly. Geoffrey Hinton, one of the researchers whose foundational work made modern AI possible, has estimated a 10 to 20 percent chance of human extinction within 30 years. Sam Altman, chief executive of OpenAI, signed a 2023 statement placing AI extinction risk in the same category as pandemics and nuclear war, and has since described his own estimate as low but not zero.

At the other end of the range, Yann LeCun, formerly Meta's chief AI scientist, puts the number close to zero. Eliezer Yudkowsky, a longtime AI safety researcher, has said it is above 95 percent, sometimes framing it closer to 99. Paul Christiano, another AI safety researcher, has put his estimate around 50 percent. Roman Yampolskiy, a computer scientist who studies AI safety specifically, has stated 99.9 percent. A 2023 survey of AI researchers more broadly found a mean estimate of 14.4 percent and a median of 5 percent, over a 100-year horizon.

There is a pattern in who lands where. Researchers working specifically on AI safety, rather than on building the systems themselves, tend toward the higher end of the range. Chief executives actively running frontier labs tend to cluster lower, in the 10 to 25 percent band, acknowledging real risk without treating it as more likely than not.

That spread is not just people disagreeing about the same question. Different estimates use different definitions of doom, extinction, permanent human disempowerment, or civilizational collapse, different time horizons, some bounded at 30 years, others at 100, others unbounded, and different conditions, the probability of catastrophe if transformative AI gets built at all, versus the unconditional probability across every possible future. Comparing Amodei's 25 percent to Yudkowsky's 99 percent as though they are answering the identical question overstates how precise any of this actually is.

The 2023 statement Altman signed was a single sentence, endorsed by hundreds of AI researchers and the heads of multiple competing labs, stating that mitigating AI extinction risk should be treated as a global priority alongside pandemics and nuclear war. It remains one of the only documents where direct competitors have publicly agreed on anything at all.

The people producing these estimates are, in most cases, also the people building the technology, and the industry built around it is now worth more than it has ever been worth. Anthropic, the company Amodei leads, was valued at 965 billion dollars following a Series H round in May 2026, and by April, secondary markets were pricing it closer to 1 trillion dollars. OpenAI sat at roughly 852 billion dollars in its most recent round. Across the ten highest-valued private AI companies, combined valuation now runs to roughly 2.37 trillion dollars, a figure that did not exist eighteen months ago in anything like its current form.

None of this is necessarily hypocrisy in the simple sense. A 25 percent chance of catastrophe also means a 75 percent chance of something else, and people who believe both numbers are real can reasonably choose to keep building, especially if they believe someone else will build it anyway, with less caution, if they stop. That argument gets made often, by people inside these companies, and it deserves to be taken at face value rather than dismissed.

We have written before about what happens to ordinary incentives once an institution's revenue stops depending on the people it serves. A related version of that logic may apply here in reverse: a lab's incentive to keep building does not actually require its own risk estimate to be low, only for the expected value of building, weighed against the risk, to still come out ahead, especially once a competitor is assumed to build regardless.

But the plain fact underneath the argument is still worth naming on its own. No earlier transformative technology was built this way, with its own chief executives publicly assigning it a real, double-digit probability of causing catastrophic harm, while simultaneously raising the money to build more of it, faster, at a valuation that assumes the good outcome. Nuclear power operators did not publish a percentage chance of meltdown while breaking ground on new reactors. This is the first one where the people closest to the technology are on the record doing both at once.