Not laid off, just not hired

Premise

Suppose that job displacement works not by pushing people out, but by keeping them from getting in.

This is a conditional, not a conclusion. But it is not made up either: it matches the pattern in what has been measured so far.

Most of today’s jobs did not exist in 1940

Whenever machines take over a task, a familiar reassurance follows: work has been displaced before, and new work turned up to replace it.

There is evidence behind that reassurance. A study that traced American job titles from 1940 to 2018 found that most of today’s employment is in job specialties introduced after 1940.

So, on that record, new work does appear.

But the same study also shows where it appeared — and that this changed over time.

Period	Where new work was created
1940–1980	Middle-paid production and clerical jobs
Since 1980	High-paid professional jobs, and to a lesser extent low-paid services

The authors report a second finding. Innovations that complement an occupation increase demand for it; innovations that automate it reduce demand. And automation’s drag on demand has grown stronger in recent decades.

Their record stops in 2018. It shows that in America’s past, new work followed displacement. It does not show that new work will follow as fast, or on the same scale, this time.

The rung that used to take people in is now the most exposed

The International Labour Organization and Poland’s national research institute have built an index of how exposed each occupation is to generative AI. A quarter of all jobs worldwide are in occupations with some degree of exposure. The share is 34% in high-income countries and 11% in low-income ones.

Clerical jobs are the most exposed. In high-income countries, the highest level of exposure covers 9.6% of women’s employment, against 3.5% of men’s.

The ILO is careful to make two points in the same document.

The figures reflect potential exposure, not actual job losses.

GenAI’s effect is more likely to transform jobs than eliminate them.

This piece goes no further than that. A quarter of jobs being exposed does not mean a quarter of jobs will disappear.

Yet put the two studies side by side and the same rung shows up in both. Between 1940 and 1980, clerical work was where new jobs were being created. After 1980, the creation of new work moved away from it. And today it is the most exposed.

In other words, the part of the job ladder that used to take in displaced workers stopped producing new work after 1980 — and it is now the most exposed.

For the argument that new work appeared before and so will appear again to hold, it matters where that work appears. The measured answer is: at the top and at the bottom. Not in the middle.

Where the change actually showed up

Payroll records from the largest payroll provider in the US, covering November 2022 to June 2026, contain a finding that has to be quoted together with its caveat.

We do not see widespread, economy-wide job displacement associated with AI.

Those are the authors’ own words. But one group in the same data stands out. Among 22-to-25-year-olds in the jobs most exposed to AI, employment is about 19% lower than it would be if it had kept pace with people the same age in less-exposed jobs. In July 2025 the gap was 15%. Among experienced workers there is no comparable gap.

How the adjustment happened is what this piece is about. In this data, it came through fewer hires rather than more people leaving, and it shows up in headcount rather than in base pay.

Those who already had jobs were less affected. Fewer of those trying to get in managed it.

The authors set out the limits themselves, and the number should not be quoted without them. These are descriptive patterns, not causal estimates; the analysis cannot establish how much of the divergence was caused by generative AI rather than by other forces; and the gaps shrink once education is taken into account.

Protection depends on how long you have already worked

Belgium’s rules put figures on this. They came into force on 1 March 2026.

Full unemployment benefit	24 months at most — a base of 12 months, plus up to 12 more depending on work history
Insertion allowance (for people starting out with no work record)	One year at most
Benefit after resigning (the so-called right to rebound)	Once in a career, for six months, and only after 3,120 days of work (about ten years)

All three point the same way. The longer you have already worked, the more protection you get.

That is not a design flaw. Linking benefits to contributions is the basic principle of social insurance, and the rule that you draw on what you have paid in has its own justification.

This piece is looking at something else. The disruption that has been measured comes before people start paying in, while protection is designed to pay out in line with what they have paid. And the March 2026 reform gave the shortest limit — one year — to the benefit for people with no work record.

So why not guarantee a minimum income?

At this point one answer tends to come up: machines will raise output, so use the proceeds to guarantee a minimum income.

To weigh that, start by asking whether there are any proceeds yet.

In 2025, 20.0% of EU businesses with ten or more employees used AI technologies, up from 13.5% a year earlier. Among large businesses the figure is 55.0%; among medium-sized ones, 30.4%; among small ones, 17%. Countries differ widely: 42.0% in Denmark, 37.8% in Finland and 35.0% in Sweden, against 5.2% in Romania and 8.4% in Poland.

What firms use it for is recorded too: analysing written text, 11.8%; generating images, video or audio, 9.5%; generating written or spoken language, 8.8%.

So have firms started to take it up? Clearly, yes: a rise of six and a half percentage points in a single year, and more than half of the largest firms.

But the productivity figures have not caught up.

In the EU, labour productivity per hour worked rose by 0.2% in 2024 and 1.4% in 2025.

And before that? Eurostat puts it plainly: average labour productivity growth slowed over 2008 to 2019, and kept slowing over 2019 to 2025. Of the three periods it reports, the strongest was 1999 to 2008.

Both can be true at once. Firms are adopting AI fast, and overall productivity has not yet registered it.

So this piece does not claim that AI has raised productivity, or that it cannot. What can be said is this: the idea that abundance has already arrived, and all that is left is to argue over how to share it, is not yet borne out by the statistics. That is not a prediction that it won’t be. It is a statement about what has been measured so far.

Labour’s share has already shrunk

If abundance has not shown up in the statistics yet, does that mean there is nothing to share? No.

The ILO finds that labour’s share of global income fell by 0.6 percentage points between 2019 and 2022, and has been flat since. Had it stayed at its 2004 level, workers would have earned US$2.4 trillion more in 2024 alone. Close to 40% of the decline happened between 2020 and 2022.

The ILO also writes that technological advances, including automation, have played a role in this trend, and that workers are not sharing equitably in the resulting gains.

So the picture looks like this. We are not in a world where productivity has soared and the question is how to split the surplus. We are in one where productivity has grown only a little and workers’ share has already fallen. Those are different problems, and they call for different remedies.

Whether those gains went to customers as lower prices or to shareholders as dividends and buybacks, we were not able to measure.

It has been put to a vote, and tested

The debate has not stayed in the abstract.

On 5 June 2016, Switzerland held a national referendum on a popular initiative for an unconditional basic income. The initiative asked that the population as a whole be able to live in dignity and take part in public life whether or not they were in paid work. The Federal Council and Parliament both recommended a no vote: 157 to 19 in the National Council, and 40 to 1 in the Council of States.

On 16 February 2017, the European Parliament adopted a resolution on civil law rules on robotics, by 396 votes to 123, with 85 abstentions. It notes that robotics may create, displace and destroy jobs, and asks the Commission to keep track of this. The text as adopted contains neither a robot tax nor a basic income. It was passed with no funding mechanism written into it.

Finland paid 2,000 unemployed people a tax-free €560 a month in 2017 and 2018, whether or not they were looking for work. The number of days they spent in work rose by six, and they were more satisfied with their lives. The ministry adds its own caveat: because a new activation model was introduced at the start of 2018, the effects in the second year cannot be separated from it.

In the United States, about 1,000 people received $1,000 a month for three years, compared with a control group of 2,000 who received $50. Labour force participation fell by 4.2 percentage points; recipients worked one to two hours less a week; and their own income, not counting the transfers, fell by about $1,900 a year. Their job quality did not improve, even though it was measured in detail, and the biggest increase in how they spent their time was in leisure. Their sense of wellbeing improved in year one, then fell back to the level of the control group. The authors conclude that there was a moderate labour supply effect that does not appear offset by other productive activities.

This piece does not decide for or against basic income on the strength of two experiments. Two are not enough for that.

In one survey, the objection was to the design

Claims about what scholars think of a guaranteed minimum income need unpacking.

In 2016, 46 leading American economists were asked about one specific design: $13,000 a year for every citizen over 21, paid for by scrapping all transfer programmes — Social Security, Medicare, Medicaid, housing subsidies, household welfare payments, and farm and corporate subsidies.

Response	Count
Strongly agree	0
Agree	1
Uncertain	6
Disagree	26
Strongly disagree	7
No opinion	6

The reasons they gave show what was being rejected. One reads: a more efficient and generous social safety net is needed, but UBI is expensive and not generous enough.

So the objection was not to guaranteeing a minimum. It was to scrapping the existing programmes and replacing them with a single payment.

One survey cannot speak for a whole field. These were 46 American economists, in 2016, asked about one design. What it shows goes this far and no further: in this survey, the objection was to the design and the financing rather than to the goal.

What is left is the cost

Being able to say no is not a matter of having the right to. It is a matter of what saying no costs you.

In Belgium, if you choose to resign, you are left with a benefit you can claim once in your working life, for six months, and only with about ten years of work behind you. Your application must reach the office within 30 days of the day after you are notified of the decision, and once it is made, it cannot be withdrawn.

So the cost depends on whether you have already been in work.

And the disruption comes before you get a foot in the door.

Where the author stands

I am on the leaving side myself, and this piece looks at people trying to get in. Those are different positions, so my own case is kept out of it.

What I can write about is the structure. In the previous piece, I asked whether the person being assessed gets to see the reasons for the assessment. This one looks at the step before that: what is left for you when you cannot even get as far as being assessed.

There is no advice here. This is as far as the measurements go.

What went wrong while writing this

Two things.

The tool I reached for at the start did not fit the question. I used an opinion poll to gauge public support for basic income, when the real question was whether the people who study where work is heading put it forward as the remedy. A poll measures what people want, not what ought to be guaranteed.

Summaries were giving a paper’s figures at half their real size. Several write-ups describe the American experiment as producing a 2-percentage-point fall in participation and 1.3 fewer hours of work a week. The paper itself says 4.2 percentage points and one to two hours. The same thing happened in the previous piece. When sources copy one another, their agreement proves nothing extra.

What this piece established

• In 2025, 20.0% of EU businesses with ten or more employees used AI, up from 13.5% a year earlier
• EU labour productivity growth slowed over 2008 to 2019, and kept slowing over 2019 to 2025
• No widespread job displacement was observed; among 22-to-25-year-olds in highly exposed jobs, employment is about 19% below that of people the same age, and the adjustment came through fewer hires
• A quarter of jobs worldwide are exposed to generative AI, clerical work most of all, and the ILO records this as exposure, not as job losses
• Most of today’s employment is in job specialties introduced after 1940, and since 1980 new work has been created at the high-paid professional end and the low-paid service end
• In Belgium, unemployment benefit lasts up to 24 months and the insertion allowance up to one year; the benefit after resigning can be claimed once in a career, for six months, and requires about ten years of work
• Of 46 economists, 33 objected to a specific design that replaced all existing transfers

What could not be checked, and was left out

• Figures from the OECD’s modelling of a budget-neutral basic income. Its website blocked access, and we did not try to get around that. We did not use the numbers circulating elsewhere instead
• The final yes and no percentages in the Swiss referendum. The official page sends readers elsewhere for the result, so only the parliamentary votes are used
• What happened to the robot tax and basic income proposals in the European Parliament’s committee. What is established is that they are not in the text as adopted

What this piece does not claim

• That AI caused the drop in youth employment. The authors say these are not causal estimates, and that the gaps shrink once education is taken into account
• That firms are trying to hire fewer people
• That basic income has failed, or that it works
• That there is a way of behaving that keeps you safe