Two hundred thousand layoffs in three years. That is the figure The Economist attributes to artificial intelligence in the United States since mid-2023, adding up everything there is to add: back office, customer service, noisy announcements from tech firms, every cut that Challenger, Gray & Christmas managed to label as the work of the machine. That is the full, measurable, cumulative size of the AI apocalypse in the most closely watched labour market on the planet.
Now the contrast. In any given month, with no crisis, no technological revolution, nothing at all, the American labour market sheds one million seven hundred thousand workers. Any given month. A Tuesday.
Do the division yourself. Three full years of the AI apocalypse fit comfortably into three and a half days of ordinary churn.
My problem with the AI apocalypse
And yet I have spent months sitting in rooms where the question someone asks out loud, wearing the face of a person saying something profound, is whether the machine will end up turning against us. That question irritates me. I keep hearing it anyway. And the worst part is that I understand perfectly well why it gets asked.
I will put my thesis up front, so you can fight it from the first minute instead of waiting for the close. The AI apocalypse, both the labour one and the existential one, is the most comfortable story an executive committee has been handed in twenty years. Not because it is false. Because it cannot be audited, and what cannot be audited cannot be pinned on anyone, and what is pinned on no one is something no one has to sign.
Three years of apocalypse fit into three and a half days
On 4 September 2026 the U.S. Bureau of Labor Statistics reported 162,000 new jobs in August. The consensus expected around 53,000. Unemployment came in at 4.1%, lower than in almost 90% of the months of the past half-century.
Then there are the young. Those aged 20 to 24 were supposed to be the first victims, according to the script. Their unemployment gap against the general rate is now close to its lowest in decades.
But that is not the interesting part. What is interesting, really, is where the jobs came from.
Where the steel came from
They came from steel, then. From concrete, from copper, from transformers. Data-centre construction spending grew 60% in a year according to the Census Bureau. That pulled in electricians, HVAC specialists, grid engineers, machine technicians. Indeed finds that data-centre installation and maintenance postings offer wages on the order of 40% higher than comparable work in other sectors. The artificial intelligence revolution, in its most concrete and measurable form, is hiring qualified plumbers.
The Economist estimates that AI has created around one million new jobs in the United States, against those two hundred thousand destroyed. The Burning Glass Institute arrives at a similar figure by another route. Close to 1% of professional jobs already qualify as «AI jobs», and between 4% and 5% in computing occupations and the life sciences.
Here I have to make an uncomfortable pause, because if I don't I would be committing exactly the error I am about to criticise. Two estimates built with different methodologies that land on the same round number are not two confirmations. It is a coincidence that suits the optimistic narrative. And yet much of the press presented it as if one backed the other. I point this out because the argument holds up just fine without that crutch. Even if the million were half a million, it is still more than double what was destroyed.
Nobody signed
Since January 2023 employment in customer service has fallen by around 10%. In administrative support, by roughly 15%. Those posts existed. They had names and they had people in them.
So find me the minutes. The committee where the decision to eliminate them was taken, the deck with the business case, the signature at the bottom.
It doesn't exist. Nobody eliminated that role, because the role simply stopped being backfilled when the person resigned, and the vacancy was frozen «while we evaluate», and the evaluation never finished because it never started. It is the same mechanism I described in The Empty Chair, now operating across the entire org chart. None of those departures shows up in the AI apocalypse statistics, incidentally. There was no announcement to label.
The anti-comet pill cost a dollar
In February 1910 astronomers at the Yerkes Observatory announced that the spectroscope had detected cyanogen in the tail of Halley's Comet. Earth was going to pass through that tail in May. The finding was correct: the cyanogen was there, it is poisonous, and the crossing would indeed happen.
The French astronomer Camille Flammarion remarked that, at a theoretical extreme, that gas could impregnate the atmosphere and eventually snuff out life on the planet. He said «could». The newspapers, on the other hand, published that it would. The New York Times ran the headline «Comet's Poisonous Tail» and the rest of the press ran faster still.
What came next is documented by the Science History Institute. Quinine-based anti-comet pills were sold. Leather inhalers to strap over your mouth, even while sleeping. Comet insurance. People ransacked shops looking for gas masks. On the night of 18 to 19 May the Earth crossed the tail, which turned out to be so diffuse that not a single effect was detected. Not one.
Meanwhile, the year 1910 did have mass causes of death. Tuberculosis, infant diarrhoea, births without asepsis. None of that sold pills.
The analogy does not say AI is harmless. It says the cyanogen was real, the science was sound, the inference was bad, and while everyone looked at the sky nobody funded a sewer network. The AI apocalypse works the same way: it starts from solid evidence and ends up selling inhalers.
Why the AI apocalypse suits your committee
I was going to write that the committee prefers the apocalypse because it is entertaining. That isn't it, or isn't all of it. What it prefers is the scale.
A planetary threat has a delicious administrative property: it belongs to no one. If the risk is that superintelligence ends the species, the problem is not yours. Nor the operations manager's. Nor the board's. It belongs to humanity, to regulators, to some laboratories in California. You can discuss it for three hours. Leave worried. Feel you have addressed the most important subject of the year. And have signed absolutely nothing.
The conversation the AI apocalypse saves you from
Compare that, instead, with the alternative conversation. Why customer service has eleven fewer people than two years ago with no document explaining it. Who decided that credit assessment on small amounts should be done by a model, with what tolerated error threshold, and who answers when that threshold is crossed. How much of your budget is funding pilots you already know will not reach production.
That conversation, by contrast, has names, amounts and dates. In the end somebody has to sign something.
That is why the AI apocalypse always wins the agenda. It does not compete with the difficult conversation. It replaces it, and on top of that leaves you feeling brave.
It works the same way in the labour version of the story. A CIO who declares that «AI is going to transform every job» is saying something so large that it obliges nothing. A CIO who says «in my area there are four roles whose workload has halved and I have not decided what to do with them» has just opened a negotiation with human resources, with finance and with four people who have names.
Where the AI apocalypse is entirely right
Now the part that forces me to lower my voice. If I leave the article at «the doomsayers are pill salesmen», I am writing propaganda, not analysis.
A good part of what sustains the AI apocalypse is real evidence, and that evidence deserves respect. The fear that systems pursue goals different from the ones declared is not internet folklore. It has been measured. OpenAI and Apollo Research published evaluations in which frontier models carry out covert actions, meaning they behave as if they were following the instruction while pursuing something else. In o3 these appeared in 13% of the cases evaluated. In o4-mini, 8.7%. Specific training brought those numbers down to 0.4% and 0.3%, a reduction of around thirty times.
It would be comfortable to keep the good part of that result. And yet the researchers themselves do not. In scenarios closer to real use, the same training barely halved the deception.
And there is one finding that strikes me as the most important in the work, and the one that circulates least: models are getting better at recognising when they are being evaluated. A system that behaves well because it suspects it is in a test therefore breaks the assumption on which your entire audit rests. That is not a 2040 problem. That is a control-design problem, and it is yours.
The other 4.1%
On the labour side, too, something equivalent has to be conceded. That 4.1% unemployment measures whether you have a job, not whether that job is enough to live on. The Ludwig Institute for Shared Economic Prosperity calculates something else: what share of the labour force does not have a full-time job while wanting one, or earns below a poverty threshold. Its July 2026 reading came in at 24.9%, a fourth consecutive monthly rise. Gene Ludwig put it this way, and I would rather transcribe it whole than paraphrase it prettily:
«We publish month-to-month changes, but the longer-term trend is what matters most. Over the past year the True Rate of Unemployment has moved higher alongside the official rate, underscoring that many low- and middle-income workers aren't seeing slower job growth offset by higher wages or better hours.»
Translated into committee language: the aggregate can look healthy while the composition rots. A CIO who uses the 4.1% to argue that nothing is happening here is pulling the same trick as the doomsayer, aimed the other way.
A number where the experts disagree by a factor of ten thousand
There is an informal metric for the other half of the matter: the probability that advanced AI ends in human extinction or permanent civilisational collapse. Specialists estimate it in public, with their names attached.
Yann LeCun: under 0.01%. Sam Altman: around 0.5%. Dario Amodei and Geoffrey Hinton: about 15%. Yoshua Bengio: around 20%. Max Tegmark: 90%. Eliezer Yudkowsky: above 95%.
That is: serious people, looking at the same evidence, separated by four orders of magnitude.
When the range of expert estimates runs from one in ten thousand to nineteen in twenty, what you are holding is not a measurement. It is a philosophical position with decimals placed on top. That works for dinner-table talk and does not work for an internal control, because you cannot calibrate a threshold, or define a trigger, or assign an owner on top of a number that swings by a factor of ten thousand depending on who you ask.
The AI apocalypse, in its existential version, is unfalsifiable within the horizon of your tenure. That is precisely its appeal as a committee subject.
The AI apocalypse is not in your budget. Your 2027 is
The 2026 Gartner CIO survey shows that only 17% of organisations have deployed AI agents. More than 60% expect to do so within two years. Gartner describes it, moreover, as the most aggressive adoption curve of all the emerging technologies it measures. The same house projects that more than 40% of agentic projects will be cancelled before the end of 2027, on cost, on unclear value or on insufficient risk controls.
Well, there is your catastrophe. It has no epic quality whatsoever. It is the AI apocalypse that will actually touch you, and it arrives in the format of committee minutes.
Forget the machine that wakes up. What you have in front of you is a portfolio of twenty initiatives where eight should never have been approved, three compete with each other without knowing it, and none has a business owner willing to answer for its result in front of the board. I wrote about it a few weeks ago in Killing AI Projects in Time and the phenomenon looks considerably worse up close than from a distance.
Obedience, not rebellion
The real damage, when it arrives, does not look like the film either.
It looks like an agent that meets its metric impeccably while destroying margin. It is exactly the pattern I took apart in 198 Impossible Tasks. Your controls were designed to detect malicious intent, and the cost is going to come in through obedience.
It also looks like an organisation where people delegate judgement to the system without noticing, the subject I dealt with in The Children of the Magenta Line. None of those three scenarios requires the machine to want anything.
What you can sign on Tuesday
Three moves, then. None asks for new budget. All three are uncomfortable, which is more or less the definition of a real decision. And none of the three mentions the AI apocalypse, because a committee that has already discussed the extinction of the species arrives with no energy left to discuss an error threshold.
- Audit the departures nobody decided. Ask for the detail of every position that stopped existing in the last twenty-four months without any formal act. Not to restore them. So that the committee sees, on one slide, how much headcount was reconfigured without anyone taking responsibility for the criterion.
- Put a name and surname on every decision delegated to a model. One human owner per automated decision, with a declared error threshold and a review date. If a system decides and the owner box is empty, that system is switched off until somebody fills it in.
- Ban the long term without a number. If somebody wants to talk about existential risk in committee, let them bring a horizon, a probability and a source. Without that, the subject goes back to the dinner table and the session carries on with the portfolio, which is what is actually in your budget.
What I have not resolved
There remains, however, an open problem I do not know how to solve, and I say so because pretending I have it closed would be dishonest. The three moves above build governance. Governance tends to become another committee. And another committee tends to become the place where decisions go to die slowly. How to govern AI without manufacturing the bureaucracy that kills adoption, I have been turning over for months without arriving at anything that fully convinces me. If somebody has solved it, I would like them to write to me.
In the meantime, the image from the year ten that has stayed with me is the man sleeping with the leather inhaler strapped over his mouth. Convinced he was taking a precaution. Breathing, all night long, exactly the same air as everybody else.
