A former Anthropic researcher, Jacob Coxon, estimated a ten per cent probability of human extinction before the end of the decade. The figure collected over a hundred million views in twenty-four hours. What it did not come with was a mechanism: no account of how the extinction would actually be carried out.
That omission is the whole story. The distance between saying the end of humanity and explaining how it would happen is an ocean, and almost nobody selling the first is willing to swim the second.
What the nuclear comparison actually teaches
Today the general public can discuss the consequences of a nuclear exchange with reasonable confidence. The extreme risk feels obvious, which makes it hard to reconstruct how poorly understood it was at the time. The Manhattan Project was top secret. Harry Truman learned of the weapons only after replacing Franklin Delano Roosevelt, and even then his information was incomplete. What was certain was that a super weapon existed, which amounted to changing the meaning of the word bomb.
Before the desert test, Enrico Fermi was not a hundred per cent certain that the detonation would not ignite a chain reaction across the atmosphere. Nothing comparable had ever been attempted. Even after two cities had been destroyed, comprehension lagged behind the fact by decades. As late as 1983 it took a made-for-television film, The Day After, watched by more than a hundred million people, to make the effects of a detonation concrete for the American public. Ronald Reagan watched it too, recorded in his diaries how deeply it affected him and how it strengthened his conviction that nuclear war must never happen, and a few years later signed the treaty with Mikhail Gorbachev eliminating entire classes of atomic weapons. Historians credit the film with a real role in that shift.
So here is the lesson, and it is not the one the apocalypse market wants quoted. Understanding a catastrophic mechanism took testing, use, decades of study and eventually a made for television drama. A percentage published on a social platform is not the beginning of that process. It is a way of skipping it.
A probability without a mechanism is a product
A number like ten per cent performs a specific function. It sounds rigorous, it cannot be checked, and it makes whoever produced it appear to be standing very close to something enormous. The person best positioned to warn you about the power of the technology is, conveniently, the person whose professional standing depends on that technology being powerful.
This is not an argument that the risks are imaginary. It is an argument that a forecast with no causal chain attached is not a contribution to safety. It is a contribution to attention.
The scenarios somebody actually worked through
The uncomfortable question has been addressed by people willing to specify. The non-profit project Future AI sketched a scenario, AI 2027, in which AI systems use a biological weapon to remove the human parasite and clear space for the solar panels needed to power their robots.
That scenario is a descendant of a familiar thought experiment: Nick Bostrom’s paperclip maximiser, from Superintelligence in 2014. Give a sufficiently capable, self-improving system a trivial objective, maximise paperclip production, and it may rationally eliminate every obstacle to that objective, including the inconvenient humans. Bostrom’s point was never a literal forecast. It illustrates two things. A system can be extremely capable without carrying any human common sense. And almost any final goal, pursued with enough power and no constraints, generates the same sub-goals: acquire resources, preserve yourself, resist being modified or switched off.
The danger, in other words, does not require the machine to be evil or conscious. It requires a gap between the objective as specified and what the humans actually meant, plus enough capability to pursue the literal version without limits. Which is a fairly good description of why learning to ask precise questions is not a productivity tip but a safety practice. We thought we were dealing with a stochastic parrot. With agentic AI we discovered flocks of millions of parrots, some of which may be cooperating with or working against each other. As usual, we had simplified a little too enthusiastically.
The RAND conclusion, and the joke inside it
More systematic work was completed in 2025 by social scientists at the RAND Corporation in California, in a study by Vermeer, Lathrop and Moon on the risk of extinction from artificial intelligence. They identify three plausible routes: nuclear weapons, a biological weapon, or the destruction of the atmosphere.
The authors dismiss the first, the most cinematic option, as not credible. In their assessment control and safety systems are already robust against technological intervention, and history offers several occasions on which catastrophe was avoided precisely because a human being refused to follow the machine. That is their judgement of a scenario, not a certificate of invulnerability, and it is worth more than a percentage precisely because they show their reasoning.
The other two scenarios are where it stops being entertaining. To deploy a biological weapon against humanity, or to wreck the atmosphere, the AI would have to use humanity. Large-scale fake news. Mass disinformation weaponry. Artificial populism. Systematic pollution of the information environment in order to put in power politicians willing to make every wrong decision on the environment while denying climate change. Destruction of common sense through algorithms engineered to turn attention into a chemical dependency. Industrial grade anti science, distributed one short video at a time.
We are the delivery system
There is one savage objection to both scenarios, and it does not come from the machines. We are already doing all of it, without any help from artificial intelligence.
Every element of that extinction plan is currently a functioning business model. Disinformation has a market. Outrage has a retention curve. Climate denial has an electoral constituency and a lobbying budget. Attention capture has quarterly targets. The supposed weapon of a hostile superintelligence is, item by item, the ordinary operating output of the platform economy, delivered enthusiastically and at no charge to the alleged aggressor.
Which leaves a conclusion nobody will put on a conference slide. If an artificial intelligence were genuinely intelligent, it would not need a biological weapon or a plan to darken the sky. It would only need patience.
Raffaele Di Marzio
All my “insane” books on cybersecurity and governance are here 👇 https://www.amazon.it/stores/author/B0FB47T6Q4/allbooks