The Economist Says Not to Blame AI for the Tech Job Collapse. For Now...

On April 13, The Economist published an article with a title that is already a statement of position: the collapse of tech jobs is real, but don't blame AI, at least not yet. In the United States, there are more than half a million tech jobs missing compared to the trajectory the sector was on before. San Francisco, the symbolic heart of the industry, has lost 3% of overall employment since the beginning of 2023. Companies like Oracle, Block, Amazon, and Meta are indeed cutting back, and the growth of headcounts among major players between 2022 and 2025 has been close to zero.

The thesis of The Economist is that the cause of all this is not artificial intelligence, or at least not yet. The more cautious explanation is a correction following the hyper-hiring of the pandemic period, when capital was cheap and companies overshot hiring, just as they did during the dot-com bubble months. Now that interest rates are higher, they are scaling back. To support this, the article cites a survey conducted on businesses in the United States, Australia, the United Kingdom, and Germany, indicating that AI has had an "essentially zero" impact on employment over the past three years. When an executive blames layoffs on AI, The Economist suggests, they are often just telling a story that sounds better to investors. I call that AI washing, attributing cuts to AI when the real reason is margin pressure.

The problem, and the reason I am writing this piece, is that "it’s not AI’s fault" and "AI is not yet capable of taking those jobs" are two different statements, and the article allows the reader to confuse them. To keep them separate, two more sources that have come out in recent weeks must be cited, which measure things The Economist does not. I will lay them out in three tiers: what AI can do, how much work it could theoretically affect, and how much it is actually affecting now.

First tier: capability, that is, what the model can do on its own The first number comes from METR, an independent group that evaluates language models. The metric they invented is called time horizon, and it’s worth explaining slowly because it’s the data around which everything else revolves. The time horizon is the length of the chain of tasks that a model completes on its own, without a human taking it by the hand, with a reliability of 50%. Translated: if a skilled human takes two hours for a certain job, and the model completes it on its own half of the time, the time horizon of that model is two hours.

This is the data I mention every time someone tells me that AI only does single tasks and not entire jobs. Because a job is held together by what I call the glue – the cost of coordinating the various pieces of work, deciding what to do first and what to do next, keeping the thread. The time horizon precisely measures how much of that glue the machine can handle on its own. When the model carries out a long sequence of steps without human involvement, it is said to work agentically, and agentic AI is exactly the novelty that is redefining the boundaries between tasks and jobs.

The METR study from March 2025 showed that this length doubles roughly every seven months, and it has been doing so for about six years. Even back then, it was a trajectory to take seriously. In January 2026, METR updated its estimates with a broader and more precise version of the test, and the picture worsened for those looking for consolation: in the period from 2024-2025, the doubling time fell from seven months to about four. In early 2026 pilot tests, the most capable models reach a time horizon estimated between sixteen and twenty hours of human work, to the point that METR warns that beyond sixteen hours, its measuring tools begin to be less reliable. That acceleration of doubling matters more than the number of hours themselves because it indicates that capacity is growing, and it's growing faster than before.

Second tier: technical potential, that is, how much work AI could theoretically affect The second number comes from a report by the McKinsey Global Institute published in May 2026, dedicated to how agents and robots are reshaping work in Europe. On one thing, McKinsey is explicit, and I give them credit for it: their main number is an estimate of technical feasibility, not a prediction of what will happen.

The number is this: in the ten European economies analyzed, which together account for more than three-quarters of the workforce and GDP of the region, 58% of current working hours could be automated with technologies that already exist. The breakdown matters: 44 points come from software agents, those that work on cognitive tasks, and 14 from physical robots. The European share is similar to that estimated for the United States, around 57%. If that potential is realized, McKinsey estimates up to $1.9 trillion in economic value could be unlocked in Europe by 2030.

Two details from the report interest me. The first: the demand for what they call AI fluency, the ability to work well with these tools, has increased fivefold since 2023, and now it appears in job postings for professions that had nothing to do with computing. The second, which is the most important and speaks directly to my thesis: three-quarters of the skills that European employers are currently looking for are needed both in automatable work and in work that is not automatable. This means that AI attacks tasks within a profession long before erasing the entire profession, and it’s the reason people are shifting towards judgment and relational tasks instead of disappearing.

Third tier: adoption, that is, how much work AI is actually touching now Here we return to The Economist, and here the source that closes the circle is added. Because if capacity accelerates and the technical potential is almost 60% of working hours, why haven’t wages and jobs collapsed yet due to AI?

The best answer I found is in a study by economists Anders Humlum and Emilie Vestergaard, published by the National Bureau of Economic Research and revised in March 2026, which examined the use of chatbots in around seven thousand workplaces in Denmark. The result, two years after the arrival of ChatGPT, is a substantially zero effect on wages and hours worked in every occupation examined. The authors manage to exclude effects greater than 2%, and report that those using these tools save, on average, 2.8% of their working hours. A crumb. The structure of work has changed, new tasks such as AI supervision have emerged, but the figure on the paycheck has remained stable.

This aligns perfectly with the “zero” from the survey cited by The Economist, and at this point, the picture becomes clearer. What is collapsing is not tech work itself. In fact, The Economist shows the most surprising data of all: between 2022 and 2025, software workers increased by 12% in retail, 75% in real estate, and almost 100% in construction. Technology is coming out of saturated major platforms and spreading into traditional sectors. The center of gravity of technical work is shifting, and an ambitious programmer today might find more space in a chain of stores than within a digital giant.

Why the three sources do not contradict each other Now reread them in order. The capability of models accelerates, and METR measures it. The technical potential for automation is enormous, around 58% of hours, and McKinsey estimates it. The actual adoption within companies, however, remains slow, with an effect on wages that is almost invisible, and that is what The Economist and the study by Humlum and Vestergaard photograph. These are three different curves, and the mistake that makes everyone relax is to confuse one for the other.

Anyone who reads only the last curve concludes that AI is a bubble of chatter and that work is safe. Anyone who reads only the first concludes that tomorrow we will all be unemployed. I believe that the honest thing to say is something else, and I say it with a metaphor that seems fitting: capacity is the power of the engine, adoption is how hard you press the accelerator. Right now, the accelerator isn’t pressed hard, and indeed the car is moving slowly. But the engine’s power continues to rise, and it is rising faster than before.

The day someone decides to press down, and McKinsey tells us that the prize for doing so is worth $1.9 trillion, the speed will no longer depend on the engine. It will only depend on how wide the margin is between where we are and where technology can already take us. And that margin today is 58% of working hours.

The Economist is right: today’s employment numbers are not enough to blame AI. But using today’s numbers to feel secure about 2027 has the same flaw as I pointed out to those citing studies with data as of 2024. The variable that matters grows after the end of the window we are looking through, and reading the final data will mean reading them when the transition has already occurred.

For those choosing what to study, or wondering if their profession will hold up, the practical conclusion is neither reassuring nor apocalyptic. It is that the right question is not “has AI already taken jobs like mine?” because the answer is almost always no today. The right question is “how large is the distance between what AI could already do with my job and what it is actually allowed to do?” The wider that distance is, the shorter your “for now” is.

It’s the same word I put in the title of my book, Humans for Now, because it tells exactly this: the form of work we know still holds for many, but it holds for now, and how long that hour lasts is what really matters.

How short your “for now” will be depends on when someone decides to press the accelerator, and I would do a disservice to anyone, myself included, if I claimed to know that with certainty.

Sources
The Economist, The tech jobs bust is real, don't blame AI (yet) (April 13, 2026).
METR, Measuring AI Ability to Complete Long Tasks (March 2025) and Time Horizon 1.1 (January 2026).
McKinsey Global Institute, Agents, robots, and us: How AI reshapes work and skills in Europe (May 2026).
A. Humlum, E. Vestergaard, Large Language Models, Small Labor Market Effects (NBER w33777).