The Fear of Being Replaced

Am I Training My Replacement?

I hear some version of that question most weeks now. It tends to come late in a conversation, and usually from someone who is good at their job. A developer who spent the afternoon steering an agent through a refactor. An infrastructure engineer who watched a script do in an hour what used to fill a sprint.

The usual reassurance is that people have always lost jobs to machines, and society adapted. That is true for the economy as a whole. It is much less true for the particular people who were displaced, and that gap is what this post is about.

This post is for developers, IT technicians and the people who hire them. It treats the fear as a labour market question you can check against evidence today. The bigger question, what work is for once we need less of it, deserves a post of its own.

Four Fears Under One Name

“Being replaced” usually bundles four separate fears. Each has a different answer, so it pays to pull them apart before looking at any evidence.

1. 1

   Income loss

   “I will be unemployed.”

   Confirmed byFalling employment in the occupation

2. 2

   Devaluation

   “The job stays, but it pays less.”

   Confirmed byWages flat or falling while output rises

3. 3

   Obsolescence

   “Fifteen years of skill stop mattering.”

   Confirmed byDisplaced workers moving to lower-paid work

4. 4

   Identity

   “I write code, and now I review what a machine wrote.”

   Confirmed byThe task mix shifts, even with stable headcount

   The most common in private, and the hardest to measure

The last one is the least discussed and, in my experience, the most common in private conversations. It is also the hardest to measure, so most of the evidence below speaks to the first three.

Society Adapted. Eventually.

“People have always lost jobs to machines” is true for the economy and false for many of the people in it. Two findings come up again and again in the history of automation.

The aggregate recovers, but slowly. The economic historian Robert Allen named the British period from about 1790 to 1840 “Engels’ pause”. Output per worker rose for decades while working-class real wages stood still, and the gains went largely to profits. Wages only caught up after the middle of the century.

The people already in the job have a different story from the people who would have joined it. Here’s the thing about the reassurance: it is built from averages, and the average hides a split.

A task is automated

• Incumbents

  Already in the job

  A decade on, more likely to be in lower-paid work or out of the labour force

  Telephone operators in post in the 1920s

• Newcomers

  Would have joined later

  Go into other work instead, and mostly do fine

  Later cohorts moved into clerical and service jobs

"Society adapted" averages these two paths over decades. Each person lives only one of them.

“Society adapted” is a statement about averages measured over decades. A 45-year-old with a mortgage does not live in an average.

A Century of Case Studies

I have picked six cases where work met a machine. In two, the fear turned out to be overstated. In four, it was justified for someone, and the interesting part is who. Step through them in order, or filter by how the fear turned out.

Fear justified, for a generation

Engels' pause

Britain, 1790 to 1840 · Steam power and the factory system

50 yearsof rising output before real wages followed

The mills will take our livelihoods.

Output per worker rose for decades while working-class real wages stood still. The gains went largely to profits, and wages only caught up after the middle of the century. Historians still argue about how flat wages really were. Nobody argues the gains arrived quickly.

Who paid

A whole generation of working families.

The lesson

Society did adapt. It took longer than a working life, which is the part the reassurance leaves out.

Sources: Allen, 2009 · Allen, working paper 315

Fear justified, for those already in post

Telephone operators

United States, 1920 to 1940 · Mechanical switching

50 to 80%fall in young women working as operators

The exchange will not need us.

Mechanical switching replaced most local operator work across more than half the US network. Later cohorts were fine: they went into clerical and service jobs, and their employment did not fall. Finishing the job took AT&T until 1978, almost 60 years, because a whole production system had been built around the operators.

Who paid

Operators already in the job. A decade on, they were more likely to be in lower-paid work or out of the labour force.

The lesson

The closest precedent for today's squeeze on junior roles. Newcomers adapt by going elsewhere. Incumbents carry the cost.

Sources: Feigenbaum & Gross, via Cato · QJE 2024 · Management Science 2024

Fear overstated

ATMs and bank tellers

United States, 1970s to 2000s · The cash machine

21 → 13tellers per urban branch, while total teller jobs held up

If a machine hands out the cash, who needs a teller?

ATMs cut the tellers an average urban branch needed, so banks opened more branches. Total teller employment held steady or rose slightly, and the job moved from handling cash to solving customers' problems. Teller numbers did fall later, because of online and mobile banking. The feared technology was not the one that cut the headcount.

Who paid

Few people, at the time.

The lesson

A job is a bundle of tasks. Automate one, and cheaper output can raise demand for the rest.

Sources: Bessen, via Conversable Economist · Knowable Magazine

Fear justified, for whole regions

The China shock

United States, 2000 to 2019 · Import competition. Trade, not automation

10+ yearsof depressed wages and participation in the hardest-hit areas

The factory is leaving, and the town goes with it.

This was trade, not automation, but it is the best-measured example of a concentrated loss of demand for labour. In the hardest-hit areas, wages and participation stayed depressed for at least a decade, and the effects were still visible in 2019. Most people adjusted by leaving work, not by moving away.

Who paid

People in one concentrated place, with no plan to help them.

The lesson

When nobody plans for a concentrated loss, people leave work rather than adapt.

Sources: Autor, Dorn & Hanson, 2016 · Autor, Dorn & Hanson, 2021

Fear overstated, so far

Radiologists

United States, 2016 to 2026 · Image recognition AI

+7%more US radiologists between 2015 and 2019

“Stop training radiologists now.” Geoffrey Hinton, 2016.

US radiologist numbers rose instead, and a shortage is forecast to deepen. By 2025 average pay had reached $571,000, up 9% on the year, and imaging demand grew, partly because AI made scans cheaper. Hinton later said he had meant image analysis only, and was wrong on timing rather than direction.

Who paid

The training pipeline. Fear alone put people off the speciality and deepened the shortage.

The lesson

The prediction confused one task with the whole job. Keep Hinton's caveat too: the direction can be right while the timing is badly wrong.

Sources: Fortune, May 2026 · AITopics · NYT, quoted

Fear justified, by market

Online freelancers

Online platforms, 2022 to 2025 · ChatGPT and its successors

−5.2%monthly earnings for writing freelancers

Clients will ask the chatbot instead.

Freelancers in the most exposed occupations lost both work and earnings. For writing work, monthly jobs fell 2%. Higher earners lost proportionately more. The effect split by market: translation shrank, while web development saw more work and higher earnings.

Who paid

Writers and translators, including the most successful ones.

The lesson

Past success offered no protection. What mattered was the market you were in, not how good you were in it.

Sources: Hui, Reshef & Zhou · arXiv 2312.04180

Where the fear was overstated, the prediction made two mistakes. It confused one task with the whole job, and it ignored how demand responds when output gets cheaper. Tellers and radiologists both did a bundle of tasks, and the machine took one of them. Hinton’s own correction is still worth keeping, though. The direction of change can be right even when the timing is badly wrong.

Where the fear was justified, the cost landed on a narrow group. It was the people already doing one specific task, or the people living in one place. That is why “the economy adapts” and “you will be fine” are separate claims. The first can be true while the second is false for you.

Developers and IT Today

The current data looks more like the telephone operators than the bank tellers. Overall demand for developers is growing, while the first rung of the ladder is shrinking. Experienced developers are mostly fine for now. Graduates are not.

• US developer employmentBLS projection to 2034 · BLS, via WebProNews
  +15%
• US developer job postingsYear on year, April 2026 · CNN, via WebProNews
  +11%
• US developers aged 22 to 25Employment, while older developers grew · Stanford AI Index 2026
  −20%
• UK graduate listingsJanuary 2026, year on year · Adzuna, via techUK
  −45%
• UK tech graduate roles2024 · ISE, via TechRadar
  −46%
• UK tech graduate rolesProjected for 2026 · ISE, via TechRadar
  −53%

Against the trendIn Denmark, across 11 exposed occupations, the effect on earnings and hours in 2023 and 2024 was statistically zero.Humlum & Vestergaard

Read with careAI was cited in 87,714 US job cuts in 2026 so far, 22% of all layoffs. A stated reason is not a measured cause.Challenger, via HCAMag

So why the juniors and not the seniors? Brynjolfsson, Chandar and Chen found that junior hiring fell where firms used AI to automate tasks, and held steady where they used it to augment their staff. James Bessen’s work points the same way from the other end: so far, cheaper software has raised demand for it rather than cut headcount. The outcome is partly a choice that each employer makes.

Two things make these numbers harder to read than they look.

A stated reason is not a cause. Forrester found that many firms announcing AI-related layoffs had no mature AI ready to replace those roles. It calls this AI washing: blaming cuts that are really financial on AI that has not arrived yet. The layoff counts probably overstate how much displacement AI is actually causing.

The baselines differ. Some reports date the fall in junior employment from late 2022, others from 2024. Higher interest rates and post-pandemic overhiring both sit inside that window, so not all of the drop is AI.

The picture for IT technicians is thinner still. I found no study for operations, support or field roles that compares with the developer research. Scripted first line support and routine operations look exposed. Hands-on hardware and field work look much less so. Treat that as my inference, not as evidence.

What Follows Depends on Where You Stand

The case studies point at different people, and each has a different problem to solve.

Experienced developers. History suggests the job changes more than it disappears, as long as demand keeps responding to cheaper output. The tasks that survived in banking and radiology were the ones bundled with judgement, context and accountability. Those are the parts of the job worth deepening. The risk to watch is your income and your sense of the craft, not your headcount.

Juniors and students. The telephone operators are your precedent: later cohorts did well by entering different jobs. In practice, that means looking for employers who use AI to augment their staff rather than automate them, and building depth in a domain early. A general junior developer role is the most exposed version of the job.

Employers. Cutting graduate intake saves money this year and creates a senior shortage in about ten. Radiology shows that fear alone can shrink a training pipeline and deepen a shortage. The outcome is partly your choice, because the firms that used AI to augment work kept hiring juniors.

Policy makers. The people at risk are incumbents in exposed tasks and a cohort of graduates who cannot get a first job, not the workforce as a whole. The China shock shows what happens when nobody plans for that: people leave work rather than adapt. Which interventions actually help is an open question, and the subject of a follow-up post.

The Work That Is Left

Here is the part of the history that rarely makes the headlines. In every case where the fear proved overstated, the work that remained was better work. Tellers stopped counting cash and started solving people’s problems. Radiologists gained software that reads scans, and demand for their judgement went up, along with their pay. Cheaper output did not shrink those professions. It created more of the work only a person could do.

Software has followed the same pattern for fifty years. Every time it got cheaper to build, from compilers to open source to the cloud, the world asked for more of it, and the number of people building it grew. AI is the biggest drop in that cost so far. If the pattern holds, more problems become worth solving than ever before: the internal tool nobody could justify, the small charity’s booking system, the process a hospital still runs on paper.

The people who gain most from that are the ones who know which problem is worth solving, and who will put their name to the answer. That was always the hard part of the job. It is now most of it.

None of this happens by itself. The operators were never given a path, and the towns hit by the China shock were never given a plan. Whether this round looks like the tellers or the operators depends on choices still being made: by employers deciding whether to hire juniors, by developers deciding what to get better at, and by governments deciding whether to plan ahead this time. For once, we can see it coming.

History’s lesson is not that you will be fine. It is that being fine is something we build, and this time we have the warning.

Appendix: evidence, confidence and sourcesHow sure I am of each claim, the biases to watch, and the research behind it.

How Sure Am I?

This is a question where a confident answer is easy to give and hard to justify, so here is where the evidence is strong and where it thins out.

• Automation's gains reach wages slowly, and incumbents often lose out

  high

  Consistent across Engels' pause, telephone operators and the China shock

• Junior developer hiring has fallen sharply in the US and UK

  high

  Several independent data sources agree

• AI is the main cause of that fall

  medium

  Interest rates, post-pandemic overhiring and AI washing all muddy the data

• Experienced developers are broadly safe for now

  medium

  Rests on demand staying responsive, from less than four years of data

• This will follow the ATMs, not the telephone operators

  low

  AI is broader than any single past technology, so the analogies may not hold

Three biases are worth naming, because they push the conversation in opposite directions.

• Survivorship bias. The developers voicing the fear are employed. The juniors who were never hired rarely appear in the conversation at all.
• Recency bias. The 2026 layoff headlines lean heavily on stated reasons, not measured causes.
• Anchoring on a kind history. The ATM story is popular partly because it is reassuring.

Most of the rigorous evidence comes from the US and Denmark, and the UK figures come mainly from employer surveys and job board data. Several of the sources below are press summaries of the underlying research, so check the original papers before you quote exact figures.

Sources

1. Allen, R. C. (2009). Engels’ pause: Technical change, capital accumulation, and inequality in the British industrial revolution. Explorations in Economic History, 46(4).
2. Bessen, J. (2015). ATMs and bank tellers, summarised in Conversable Economist.
3. Bessen, J. (2026). Why AI hasn’t killed software developer jobs. TPRI Report, Boston University.
4. Feigenbaum, J. & Gross, D. P. (2024). Answering the Call of Automation. Quarterly Journal of Economics, 139(3).
5. Feigenbaum, J. & Gross, D. P. (2024). Organizational and Economic Obstacles to Automation. Management Science, 70(12).
6. Hui, X., Reshef, O. & Zhou, L. Short-term effects of generative AI on freelancer employment. Organization Science.
7. Autor, D., Dorn, D. & Hanson, G. (2016). The China Shock. Annual Review of Economics, 8.
8. Autor, D., Dorn, D. & Hanson, G. (2021). On the Persistence of the China Shock. Brookings Papers on Economic Activity.
9. Humlum, A. & Vestergaard, E. (2025). Large Language Models, Small Labor Market Effects. NBER Working Paper 33777.
10. Brynjolfsson, E., Chandar, B. & Chen, R. (2025). Young workers in AI-exposed occupations, summary.
11. Stanford HAI (2026). AI Index Report 2026.
12. Institute of Student Employers, UK tech graduate hiring, via TechRadar.
13. Challenger, Gray & Christmas and Forrester on AI-cited layoffs, via HCAMag.
14. Hinton and radiology, via Fortune.