Seven in ten expect job losses. That is not what we measure.

Pew Research Center asked 42,151 people across 36 countries, plus two separate surveys in the United States, what artificial intelligence will do to work. The answers came back with unusual agreement.

In 34 of the 37 countries in the study, people lean towards AI meaning fewer jobs rather than more. In Australia, South Korea and the United States, around seven in ten adults or more say AI will cost jobs over the next twenty years.

That is the clearest public verdict yet on the question this site exists to address. It is also a different question from the one this site answers, and the difference is worth being exact about.

What the survey measured

Pew measured what people expect, on a twenty-year horizon, about jobs in general. Not their own job. Not this year. Expectations.

Expecting a change is not the same as dreading it, and the survey keeps the two apart. Across the 37 countries, a median of 37% say they are more concerned than excited about AI, while a median of 41% say they feel both about equally. Israel is the only country where more people are primarily excited than primarily concerned. South Koreans are among the most likely to predict both job losses and widening inequality — and among the least likely to describe themselves as primarily concerned about the technology.

So the headline is not that the world is alarmed. It is that the world has settled on a direction.

What this site measures

Every score here starts from tasks. The International Labour Organization broke 427 occupations into 3,265 everyday tasks and had each one rated for how much of it today’s AI tools could do. A score on this site answers one question: of the tasks in this job, how many are the kind a language model is good at?

It answers that as of the day the ratings were published. It has no view on headcount in 2046. It does not model what an employer will do with the time a tool saves, what a tool costs to run, what wages do, or what new work appears. Those are the steps between “these tasks are exposed” and “there will be fewer of these jobs”, and this site does not take them. How it works says so in more detail than most readers want.

Which is why a customer service rep scores low here and a plumber scores high, and neither number tells you whether anybody gets hired.

Where the two findings sit next to each other

There is a real tension here, and it is worth naming carefully rather than resolving it cheaply.

The one source in our register that separates augmentation from automation — the Jobs and Skills Australia capacity study — found 349 of 357 Australian occupations to be augmentation-led. Read quickly, that sounds like the opposite of what seven in ten Australians told Pew.

It isn’t, because the two are answers to different questions. “This occupation is augmentation-led” is a statement about the shape of the work. “There will be fewer jobs” is a statement about totals, and totals depend on decisions nobody has measured yet: whether a firm takes the time saved as more output or as fewer staff, and whether the demand exists to absorb the output.

An economy can be augmentation-led occupation by occupation and still employ fewer people. The task data cannot rule that out. It cannot support it either. The public may well be right; nothing on this site can confirm it, and anyone telling you the exposure numbers refute the survey is reading both of them wrong.

The people who have to act on it first

The sharpest movement in the survey is generational. Since 2024, the share of Americans aged 18 to 34 who are more concerned than excited about AI has gone from 40% to 55%. That is a fifteen-point shift in two years among the people currently choosing courses, apprenticeships and first jobs.

This matters beyond sentiment, because a belief acts on the world before a measurement does. Applications to a course fall, and five years later there are fewer graduates — whether or not the forecast that caused the fall was any good. Hiring managers who expect a role to shrink hire fewer people into it, and the role shrinks. Expectations are not a forecast of the labour market so much as one of its inputs.

That cuts both ways, and it is the reason a page like work that is forming exists: the roles that are actually hiring right now are mostly ones nobody’s expectations had a name for three years ago.

What to take from it

Two things, held separately.

If you want to know what people think will happen to jobs, read the survey. It is the best-measured answer available, across 37 countries, and it is remarkably consistent.

If you want to know what today’s tools can already do with the tasks that make up your week, that is the smaller and duller question this site can answer. Find your job, then compare its task list against what you actually did yesterday.

One is evidence about the future. The other is evidence about the present. Only the second one is a measurement, and the first one is more likely to be what changes your workplace next year.