Technology has always created new jobs. Will it this time?

Two reports came out a month apart this year. Stanford researchers went through American payroll records and found no sign of AI wiping out jobs across the economy. But people aged 22 to 25 in the jobs most exposed to AI now have employment 19% below where it would be if it had kept pace with people their age in less exposed jobs. A year ago that gap was 15%.

The International Federation of Robotics, the robot industry's trade body, reports that factories installed more than 600,000 new robots in 2025. China took 59% of them.

So is this time different? This first part of a series on technology and everyone's work shows what the answer depends on. That thing started slipping almost forty years ago.

We've heard this before

The track record of people who predicted the end of work is terrible. In 1995, a book quoted a union leader predicting that within thirty years, as little as 2% of the world's workers would be needed to make everything we want. Those thirty years ran out last year.

But "it has always worked out" is an observation, not an explanation. A turkey fed every morning has a thousand days of evidence that mornings are safe, right up until Christmas.

Machines take over tasks, but technology also creates work in two ways: things get cheaper, so people buy more, and entirely new kinds of work appear. About half of US job growth from 1980 to 2015 came in occupations whose titles or tasks had changed.

So it's a race between work disappearing and new work appearing. Daron Acemoglu, who shared the 2024 Nobel Prize in economics, and Pascual Restrepo say the assumption that all technology raises the demand for workers "is wrong".

They tried to measure that race for the United States. From 1947 to 1987, machines took away about half a percent of the demand for workers each year, and new work added almost exactly that back. From 1987 to 2017, machines took work away about twice as fast as new work appeared.

These are estimates, for one country. But this happened before AI could write a sentence.

Where do you go next?

Every earlier wave of machines left an exit. Those of the Industrial Revolution, wrote Joel Mokyr, who shared the 2025 Nobel Prize, and two co-authors in 2015, "could not reason, compare, compute, read, smell, sense, hear, or make snap decisions". If future machines could, "it seems frighteningly plausible that this time will be different".

Mokyr has since come down on the other side. Asked in 2025 whether this time is different, he said "the answer is basically: No", and that the coming shortage is of workers, not jobs. I find that hard to share: a shortage only protects jobs if machines can't fill it.

That exit is narrowing. On software engineering tests, AI's score went from 4.4% to 71.7% between 2023 and 2024. The International Labour Organization estimated in 2025 that about one job in four worldwide has some exposure to generative AI, though that's an upper limit and it expects most of those jobs to change rather than vanish.

If work that needs a head isn't safe, the way out is work that needs hands.

The robots are going to China

China installed 354,200 industrial robots in 2025, against 38,400 in the United States. And it's far from finished: with 166 robots per 10,000 manufacturing workers (2024), it ranks 22nd in the world. South Korea has 1,220.

What do robots do to the people working next to them? Three countries give three answers:

• China. From 2010 to 2016, one standard deviation more exposure to robots went with 7% lower employment and 8% lower hourly earnings.
• United States. From 1990 to 2007, one more robot per thousand workers lowered the share of local people in work by 0.2 percentage points and wages by 0.42%.
• Germany. No net loss. New service jobs made up for the factory jobs that went, and the cost fell mostly on young people just entering the workforce.

So the robot alone doesn't decide the outcome. Nor does every robot replace someone: in ageing countries, one can fill a gap as workers get scarce.

The next robot has legs

All of these studies measure a machine bolted to the floor. It stayed in factories, so the people it displaced could move elsewhere, in Germany into services. A machine shaped like a person could, in principle, work in warehouses, shops and care homes as they already are, and follow people there.

I couldn't find a study of what humanoids do to jobs, because so few have been at work long enough. So this section rests on market counts, forecasts and press reporting.

Humanoids are new, so the change says more than the total. Smart Analytics Global counted about 19,100 shipped worldwide in the first half of 2026, almost four times the 5,100 of a year earlier, and projects close to 60,000 for the year. Estimates for all of 2025 run from about 13,000 to about 20,000. The counts differ, but they all point the same way: the research firm GGII, looking at China alone, expects shipments to nearly triple from 18,000 to more than 50,000.

Those numbers are still small next to more than 600,000 new industrial robots a year. But Bank of America forecasts 1.2 million humanoids a year by 2030, more than the 806,000 industrial robots the trade body expects for 2029. And in the first half of 2026, China's governments spent at least $230 million on humanoids and related items, up from $62 million a year earlier.

Startups now pay people to wear cameras while they work, so robots have something to learn from, some for around US$3 an hour. It's new work whose whole purpose is to make other work unnecessary.

For now, humanoids aren't very good. At a government-funded training centre in southern China, one staffer told Reuters the robots complete simple tasks at "just 20% of a human's speed or output". GGII expects at least 15% of China's humanoid makers to fail or be absorbed within three years. This could still end as a bubble.

But the consultancy Roland Berger models running costs of around two dollars an hour "in the future", without saying when. As an illustration only, a robot at that cost and a fifth of a person's speed matches a person paid ten dollars an hour, before supervision, breakdowns and mistakes. A humanoid doesn't have to be as good as a person, only cheap enough per finished task.

If humanoids succeed, the factory studies stop being a guide. So where do people go next?

Nobody knows, and that cuts both ways. David Autor, a leading labour economist, argues that automation also complements workers and can increase the demand for labour. In the Stanford data, employment is flat or rising where AI mostly helps people rather than replaces them.

The machines may also turn out less capable than promised. That wouldn't settle it either.

What the experts actually disagree about

Some economists doubt that AI will change much soon. William Nordhaus, who shared the 2018 Nobel Prize, ran seven tests on whether technology is about to send growth through the roof. Five said no, though his data ends before today's AI. Acemoglu estimates AI will add no more than about 0.66% to productivity over ten years.

That isn't reassuring. A phone menu replaced someone without making anything much cheaper, so it created little new work. Acemoglu and Restrepo call this "so-so" technology: the threat is the mediocre machine that's just good enough to replace you.

And if AI does improve fast? A large survey from 2026 asked economists and people working in AI what AI will do to the US economy at different speeds of progress. If it improves rapidly, economists put 2050 unemployment at about 6%, barely above the 5% they expect anyway. People working in AI and AI policy said 8%.

Speed isn't the main divide. The disagreement comes mainly from different beliefs about what capable AI does to the economy once it exists. As I read it, that's the race again.

Who is feeling it first

In the Stanford study, there's no wave of layoffs. The gap comes mainly from less hiring, and nobody writes a news story about a job that never appeared. It's the same pattern as the robots in Germany.

But the authors stress that their numbers don't prove AI caused it. Euro area youth employment in information and communication fell 18.6% between early 2023 and early 2026, and European Central Bank staff write that the evidence "does not yet allow this pattern to be attributed to generative AI or remote work specifically".

Where that leaves us

It hasn't been proven that this time is different, and it can't be dismissed. History rests on new work appearing as fast as old work disappears, and in the US that stopped around 1987. Now AI is moving into mental work and humanoids are being built to follow it into physical work.

Five signs would tell you the wave has arrived:

1. The gap spreads from the youngest workers to older ones.
2. Unemployment rises across whole economies, in more than one country.
3. Workers' share of income falls.
4. Humanoids do paid work at close to human speed outside factories.
5. New work falls further behind the work that disappears.

Factories wiped out Britain's handloom weavers quite quickly after 1815. Economic historians call that "by and large a one-generation affair". If you're in that generation, it's your whole working life.

So who is in that generation this time, and what happens to them when their income has gone but the rent hasn't? That's the next part.