Can AI predict the end of the world?

The number that made the debate viral again

The latest flare-up in the AI-risk debate began with a stark claim: Alex Turner, a former Google DeepMind alignment researcher, has estimated a 25–30% probability that AI contributes to the death of at least one billion people by 2050 [1]. In the French tech discussion that pushed the story back into circulation this week, the central point was not only the size of that number, but the machinery that produces such forecasts: expert judgment, lab culture, political anxiety and the accelerating deployment of powerful systems [2].

Turner’s framing matters because it is narrower and more concrete than the usual “AI extinction” slogan. He is not simply saying that a malicious superintelligence wakes up and decides to wipe out humanity. The version summarized by Silicon Carne and related podcast listings places more weight on misuse: governments, militaries and security agencies obtaining increasingly capable AI systems and using them in destabilizing, coercive or badly controlled ways [1].

That distinction changes the debate. If the risk is mainly “rogue machine,” then the problem appears technical: alignment, interpretability, control. If the risk is mainly “humans using very capable systems badly,” then the problem is institutional: contracts, incentives, oversight, military doctrine and political leadership.

What P(doom) does — and does not — measure

“P(doom)” is a shorthand used in parts of the AI-safety world for the probability of catastrophic outcomes from advanced AI. In Turner’s case, the threshold being discussed is not the literal end of all life on Earth, but the death of at least one billion people before 2050 [1]. That is already extreme enough to justify serious attention, but it is also a reminder that different people use “doom” to mean different things.

The weakness of P(doom) is that it can sound more precise than it is. A number like 25%, 30% or 10% carries the aura of measurement, yet there is no repeatable experiment for “AI causes a billion deaths by 2050.” There are no historical base rates for artificial general intelligence, no settled definition of superintelligence, and no agreed model that converts today’s benchmarks into geopolitical casualty estimates. The HelloBro summary of the episode captures that skepticism directly: critics see these figures as subjective forecasts rather than reproducible scientific measurements [4].

Still, subjective does not mean useless. Weather forecasts, election models and pandemic scenarios all include judgment calls. The difference is that those domains have feedback loops. AI-catastrophe forecasts do not: by definition, the worst cases happen once or not at all. That makes P(doom) less like an engineering metric and more like a public-risk signal. Its value is not precision; its value is whether it forces better questions.

Turner’s departure and the military thread

A key detail in the current discussion is that Turner’s departure from Google DeepMind is being linked not just to abstract existential risk, but to concern over government and military uses of AI. The podcast transcript surfaced by Podscan says Turner came from the AI-safety world, worked on alignment at Google DeepMind, and left after objecting to government contracts for military AI use that he believed lacked sufficient ethical safeguards [3].

That is why the story resonates beyond the usual alignment community. The immediate fear is not only a future supermind escaping a lab. It is the much more ordinary possibility that powerful tools become embedded inside defense and intelligence systems before democratic oversight, technical evaluation and legal accountability catch up.

This is also where the “end of the world” framing becomes misleading. The nearer-term concern is not a cinematic apocalypse; it is delegation. Which decisions get handed to AI agents? Who audits them? What happens when an autonomous system optimizes a military objective too aggressively? Who is responsible when a model-generated assessment escalates a crisis?

Those questions are not theoretical theater. Fresh policy commentary this week argued that AI-enabled adaptive computer worms could become a concrete national-security crisis, precisely because they could revise attack strategies as they spread and place sophisticated cyber capabilities within reach of non-state actors [5]. That does not prove Turner’s 25–30% estimate. It does show why misuse scenarios are not limited to science fiction.

Why the number irritates critics

The strongest criticism of P(doom) is not that catastrophe is impossible. It is that probabilistic doom-talk can smuggle ideology into what looks like quantitative analysis. If a researcher says “30%,” the audience may hear a scientific conclusion. But the estimate may encode assumptions about timelines, geopolitics, human incompetence, corporate secrecy, military escalation and whether AI capabilities will continue scaling.

Silicon Carne’s episode leans into that discomfort, asking whether these probabilities are “as reliable as a poker bluff” and whether a subjective estimate has been elevated into a kind of truth [2]. That is a useful provocation. If a number cannot be independently tested, journalists, policymakers and readers should ask what work the number is doing. Is it clarifying uncertainty? Attracting attention? Supporting regulation? Justifying a pause? Warning against military deployment? Or marketing the importance of the people and labs producing the warning?

There is also a cultural critique. Large AI labs recruit highly technical people to build systems of enormous commercial value, then ask them to reason about civilization-scale harms under secrecy, competitive pressure and moral stress. The HelloBro summary highlights the possibility that lab culture itself can amplify apocalyptic thinking: researchers arrive expecting open science and encounter product urgency, internal conflict and the sense that their work may shape civilization [4].

That does not invalidate their warnings. It does mean the warnings should be read with context.

The other side: dismissing the risk is also lazy

If P(doom) can be overconfident, pure dismissal can be worse. A precise percentage may be questionable, but the underlying risk channels are real enough to deserve governance: autonomous cyber operations, AI-assisted weapons targeting, mass surveillance, biosecurity misuse, model deception, dependency on opaque systems and concentration of capability in a few corporate or state actors.

Turner’s estimate is controversial because it collapses many of those pathways into one headline number. But the controversy should not distract from the practical agenda. Even if the true probability were far lower than 25–30%, a billion-death scenario is so severe that society should not treat it as a normal product-risk category.

The right response is not panic; it is decomposition. Separate technical alignment from misuse. Separate extinction from mass casualty risk. Separate model capability from deployment context. Separate what labs can test internally from what governments must regulate externally. Once the problem is decomposed, the debate becomes less mystical and more actionable.

Can AI predict the end of the world?

Not in the way people often imagine. AI cannot currently compute a reliable date for its own apocalypse, and humans cannot validate a neat probability for unprecedented civilizational failure. But AI can help model components of risk: cyber propagation, supply-chain fragility, command-and-control failures, biological-design workflows, infrastructure dependencies and escalation pathways.

The harder question is whether the people and institutions building AI can predict themselves. Can labs forecast how their tools will be used once they are integrated into states and markets? Can governments resist deploying systems before they understand them? Can researchers distinguish between a warning, a worldview and a measurement?

Turner’s 25–30% estimate should therefore be treated neither as prophecy nor as nonsense. It is a stress test for public reasoning. If the number shocks us into asking who controls frontier AI, how military uses are constrained, and why catastrophic forecasts are so hard to audit, it has served a purpose. If it becomes just another viral doom statistic, it has failed.

The most honest answer to the title question is this: AI cannot predict the end of the world. But the debate around AI may reveal how bad humans are at predicting the consequences of power before they deploy it.

Sources from the last 72 hours

1. [1]Le risque existentiel de l’IA est-il sérieux ?Sep 22, 2026, 12:00 AM UTC
2. [2]Peut-on prédire la fin du monde avec l’IA ? | Silicon Carne, un peu de picante dans un monde de Tech ! | AushaSep 22, 2026, 12:00 AM UTC
3. [3]Silicon Carne, un peu de picante dans un monde de Tech ! - Podcast Analytics & Insights - Podscan.fmSep 22, 2026, 4:00 PM UTC
4. [4]Peut-on prédire la fin du monde avec l’IA ? · IA · HelloBro.aiSep 22, 2026, 4:00 PM UTC
5. [5]Trump and Xi Should Talk About WormsSep 22, 2026, 12:00 AM UTC

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.