Work

A job disappears fast. The one that replaces it takes years to arrive.

'New jobs will come.' Sure. Just not in time for the people losing theirs now.

Whenever AI and jobs come up, the same line follows almost every time: don’t worry, new jobs have always come along. And that isn’t even wrong. Around 60 per cent of the jobs that exist today did not exist eighty years ago.

It just answers the wrong question. The point isn’t whether new work appears, but when. And that is exactly where the two come apart: work disappears faster than ever, while new work grows back as slowly as it always has.

A technology spreading this fast is the new part. It used to take decades, because it was tied to factories, machines and power lines that had to be built first. Software needs none of that. It is distributed once, and that’s it. The share of EU enterprises using AI rose from 8 to 20 per cent in just two years.

The growth of new work can’t be sped up the same way. New jobs don’t simply fall out of the technology. They appear only once new industries grow, once people learn new skills, once whole training paths are in place. That takes time. Those 60 per cent of new jobs built up over eighty years, not overnight. Anyone starting out young in an AI-adjacent job today feels the fast side first.

That such a gap stays open for years is no tale from the 19th century, even if the last great wave ran exactly the same way. When China pushed entire US regions out of the market from the 1990s on, wages there were still on the floor a decade later, and the new jobs in other industries weren’t there. The upturn did come in the end. It’s just that more than ten years sat in between, years in which the people on the ground simply lost out.

None of this is inevitable. ATMs were seen as the end of the bank teller, and yet for years there were more of them, because branches got cheaper and banks simply opened more. How big the gap gets and how long it stays open does not depend on the technology. It depends on whether we prepare for the time in between, or just hope it sorts itself out. That things get better in the end may well be true. It’s just that nobody plans for the years until then.

We need to talk about this

What does this topic stir up in you?

In depth

Disappearing is faster now

The economists Diego Comin and Bart Hobijn measured the spread of 15 technologies across 166 countries and almost two centuries (American Economic Review, 2010). On average it took 45 years for an invention to become widely adopted. The reason was almost always the same: a new technology was tied to factories, machines, power lines, to capital that had to be built first.

Software is tied to none of that. According to Eurostat, the share of EU enterprises with at least ten employees using AI rose from 8 per cent (2023) to 13.5 (2024) to 20 per cent (2025). In Austria, according to Statistics Austria, it nearly doubled within a single year, from 10.8 to 20.3 per cent. These figures only say how fast the technology spreads, not how many jobs it costs. But that is the point: what used to be the building of a factory is now an update.

Measuring the two sides separately

Daron Acemoglu (Nobel Prize in Economics 2024) and Pascual Restrepo set out to measure both forces separately for the US (Journal of Economic Perspectives, 2019): the disappearance of tasks to machines and the emergence of new ones. From 1947 to 1987 the two were almost exactly in balance. One lowered demand for labour by 0.48 per cent a year, the other raised it by 0.47.

From 1987 to 2017 that tipped: the disappearance accelerated to 0.7 per cent a year, the emergence fell back to 0.35. What is measured here is the task content of work, not the net number of jobs. But the direction is clear. Since the late 1980s, work has been disappearing faster than new work arrives.

Why the new jobs don't arrive on their own

That new work emerges is well documented. David Autor and colleagues tracked some 35,000 job titles since 1940 (New Frontiers, Quarterly Journal of Economics, 2024): about 60 per cent of US employment in 2018 sat in job profiles that did not exist in 1940, and among professionals as much as 74 per cent.

The decisive finding of the same study: this new work does not grow out of automation itself. It appears where technology complements people's work rather than replacing it, and that is slower. The 60 per cent built up over eight decades. How to bring that up to the speed of a software rollout, nobody has shown to this day.

A lost decade, and a recent one

You don't need the 19th century for this. When China entered world trade from the 1990s on, entire US regions lost their industry. David Autor, David Dorn and Gordon Hanson show (Annual Review of Economics, 2016): wages and employment there stayed depressed for at least a full decade, and the new jobs in other industries had not yet appeared.

A model of the same shock (Caliendo, Dvorkin, Parro, Econometrica, 2019) puts a number on the gap: shortly after the shock the net welfare gain for the US was close to zero, at 0.2 per cent. The large gain of 6.7 per cent came only once people had moved into growing industries and regions over roughly ten years. The gain was real. It just came a decade too late for those who lost first.

Displacement is no law of nature

The fast side can be slowed down too. In the 1990s the ATM was seen as the certain end of the bank teller. The economist James Bessen did the maths (2015): tellers per branch fell from 20 to 13 (1988 to 2004). But because branches became cheaper, banks opened 43 per cent more of them. Despite more than 400,000 machines, the total number of bank tellers held steady at first.

That is the honest other side: when a task gets cheaper, it can actually expand demand, and with it employment, rather than destroying it. How big the gap gets and how long it stays open does not depend on the technology. It depends on what we do during that time.

Sources