Skip to content
Go back

Jobs of the Future: 60% of Today's Didn't Exist in 1940

Illustration of a future horizon with icons of growing professions

In 2024 David Autor, Caroline Chin, Anna Salomons and Bryan Seegmiller published in the Quarterly Journal of Economics a study that required reconstructing, title by title, every occupational designation recorded in the United States from 1940 to 2018. They used occupational dictionaries and patents to date the appearance of each job.

The headline result: about 60% of 2018 employment was in occupational titles that did not exist in 1940.

That number says two things at once, and it matters not to keep only one. The first is that work renews itself continuously, and that the mass unemployment predicted at every technological wave has never materialised. The second, less comforting for list-writers, is that the most important occupations twenty years from now have no name today, and so cannot appear on any list.

Why occupational forecasts go wrong

There’s a way to check this rather than merely suspect it.

The US Bureau of Labor Statistics publishes ten-year occupational projections and, rarely and admirably, also publishes evaluations of its own past errors. Its self-assessments show a consistent pattern: aggregate projections of total employment are reasonable, while error grows considerably as you descend to the individual occupation, where forecasts are often wrong even in the direction of change.

The general picture on long-range forecasting comes from Philip Tetlock, who between 1984 and 2004 collected roughly 28,000 predictions from 284 experts on political and economic events, and scored them once the events occurred. The result, published in 2005, became famous: expert forecasts were barely better than chance, and the most confident and most media-present experts were systematically the least accurate.

That isn’t an argument against thinking about the future. It’s an argument against believing specific lists.

What can be predicted well: tasks

There’s a strand of economic research that works much better, and the difference lies in the unit of analysis.

In 2003 David Autor, Frank Levy and Richard Murnane proposed a framework that still holds: technology doesn’t replace occupations, it replaces tasks. And the distinction that matters isn’t manual versus intellectual, it’s routine versus non-routine. A routine task is one whose rules can be written out in full, whether that’s tightening a bolt or filling in a form to fixed criteria.

The prediction that follows has been tested at scale. Maarten Goos, Alan Manning and Anna Salomons documented in 2014, across 16 European countries, the phenomenon of polarisation: employment grows at the high-skill end and at the low-skill but hard-to-automate end, while the middle band of administrative clerks, line operators and bookkeepers hollows out. Autor and Dorn had found the same pattern in the United States.

That prediction, made in the early 2000s, came true. The lists of the most promising jobs from the same years did not.

The ATM case

The most useful example of why “this job will be automated” is almost always the wrong question comes from James Bessen.

The cash machine looked like the end of the bank teller. Bessen reconstructed the data: the number of ATMs in the United States went from around 100,000 in the mid-1990s to over 400,000 by 2010. Over the same period the number of bank tellers did not fall. It rose slightly.

The mechanism: ATMs lowered the cost of running a branch, banks opened more of them, and every branch needed staff. The teller’s job did change, though: less cash counting, more product sales and customer relationships.

That’s the general model. Automating a task changes an occupation’s content and its relative price long before making it disappear. And it’s why the gap between Frey and Osborne’s estimate, which put 47% of US jobs at risk, and the OECD’s, which put it at 9%, depends almost entirely on the unit of measurement: whole occupations versus tasks within occupations. I wrote about that in detail in will AI replace work?.

Nobody can tell you what job you’ll do. But it can be said with reasonable confidence which parts of your current job will go first.

The directions the data supports

With all the caution the rest of this article demands, these are the directions supported by structural data rather than industry forecasts.

DirectionWhat it rests onConfidence
Care and personal assistanceDemography: age structure already observed, not forecastHigh
On-site technical maintenance and installationNon-routine, non-offshorable tasksHigh
Energy transition and building retrofitRegulatory commitments and investment already decidedMedium-high
Working with data and interpreting itDocumented growth in demand for analytical tasksMedium
Information security and complianceDemand driven by regulation and attack surfaceMedium
Coordination roles between people and systemsExtrapolation, less solidLow-medium

The demographic row is the one you can count on most, for a mundane reason: the people who will be eighty in 2050 have already been born and already counted.

How to prepare without guessing the future

  1. Analyse your job by task, not by title. Write down the ten things you do in a week and mark which follow rules that can be written out in full. That list is your real exposure.
  2. Don’t chase the fashionable occupation. By the time a profession makes the lists, the window of advantage has largely closed and the cost of entry is at its peak.
  3. Build combinations, not pure specialisations. Hard-to-fill positions almost always arise at the intersection of two fields, not at the summit of one.
  4. Aim at non-offshorable, non-routine tasks. Physical presence, legal responsibility, negotiation, judgement under ambiguity.
  5. Account for the real cost of switching. Changing sector isn’t free, and the figures on the size of the income loss are detailed in reskilling.
  6. Prefer skills that transfer. The useful distinction isn’t hard versus soft: it’s between what’s worth something only inside one company and what’s worth something anywhere. The point is developed in hard vs soft skills.
  7. Reconsider every twelve months. Not because the world changes every year, but because it’s the only way to notice early when it does.

My take: the lists do more harm than good

I think lists of the jobs of the future do more harm than good, because they shift attention from the useful question (which parts of my job are replaceable) to a question nobody can answer.

I also think the Autor et al. 60% figure is almost always cited by halves. The optimistic part is true: jobs we can’t imagine today will appear. The omitted part is that, historically, the people who lose the old job and the people who get the new one are often not the same people, or even the same generation.

My sense of it, which remains a hypothesis: the rational move under high uncertainty isn’t betting on the right sector, it’s reducing the cost of being wrong. That means transferable skills, a network outside your own industry, and enough financial margin to change without having to do it in a hurry.

Frequently asked questions

Which jobs are safest? None is safe in absolute terms. The most automation-resistant, by task structure, are those combining non-standardised physical presence, responsibility and judgement in ambiguous situations.

Will AI destroy more jobs than it creates? Historically that has never happened, but the past isn’t a guarantee. The important point is that the net effect and the effect on individual people are two different things.

Is it worth studying computer science now? Demand for analytical tasks stays high, but the trade is changing fast. It’s worth it if your interest is genuine; as a purely strategic bet it’s riskier than it was ten years ago.

What happens to the middle band? It’s the one documented as most exposed by the research on polarisation: administrative and operational roles with writable rules. It’s also where retraining early pays off most.

How do I choose without knowing the future? By choosing criteria instead of destinations: non-routine tasks, transferable skills, sectors where demand rests on already-observed facts such as demography.

Sources

This article was written following the process described in how we write. If you find a figure reported wrongly, write to me: I’ll correct it and note the correction at the bottom of the article.

Founder of UpSkillLab, with a Master's in Digital Marketing from the University of Florence. Every article starts from public sources cited in the text and is drafted with the help of AI tools, under my review and editorial responsibility: no purchased content, nothing published without human checking. Read the editorial method or report an error.


Share this post:

See all articles in Industries & Trends

Previous Post
Green Jobs: Research Measures Them at 2-3% of Employment, Not Millions
Next Post
Microsoft Copilot: What the Controlled Trials Say (and Don't)