Job Displacement and Creation

In the spring of 2026, a software engineer we'll call Marcus refreshed a job board for the ninety-somethingth time and noticed something that hadn't been true a year earlier. The listings his skills matched—junior backend roles, entry-level full-stack positions, the rungs he'd climbed himself half a decade ago—had thinned to almost nothing. His own layoff, part of a tech-sector purge that would eventually claim more than 139,000 positions in the first half of 2026 alone (Challenger, Gray & Christmas, 2026), had not been the strange part. The strange part was that the ladder beneath him seemed to have been pulled up. Nobody was hiring the person he had been at twenty-three.

This is the shape of the thing we need to understand. Not a sudden collapse—the headlines promising that never quite arrive—but a quiet reorganization of who gets to work, at what level, for how much, and whether the first job that once led to the second one still exists at all. To see it clearly, we have to hold two facts in our hands at the same time, and resist the urge to drop either. AI is destroying real livelihoods right now. AI is also creating real work right now. The interesting questions live in the space between those two facts: who wins, who loses, how fast, and whether the people losing are the same people winning.

Echoes, and Where They Break Down

We reach instinctively for history, because we've watched machines eat jobs before.

In 1764, a British carpenter named James Hargreaves built a wooden frame that could spin eight threads at once, then sixteen, then more. His Spinning Jenny meant one person could do the work of many, and within a few decades the skilled handloom weavers of England watched their apprenticeships turn worthless. They did not go quietly. Between 1811 and 1816, textile workers who called themselves Luddites smashed the machines that had taken their trade. History remembers them as fools who stood in front of progress, but that verdict is unfair. Roughly 50,000 textile workers lost their positions, and those who found factory work earned a fraction of what the artisans before them had. The weavers were not wrong to be afraid. They were only wrong about whether fear would save them.

The pattern repeated. Steam railroads erased perhaps 200,000 jobs in horse-drawn transport across Europe and North America. The computerization of data processing in the 1960s and 70s wiped out around 400,000 clerical positions in the United States—typists, filing clerks, bookkeepers, the respectable middle-class jobs of their day. Each time, the economy eventually generated more work than it destroyed. Each time, the transition was long and cruel to the people caught inside it.

So the comfortable story goes: this is creative destruction, we've survived it before, we'll survive it again. There is truth in that. But two features of the current wave break the old analogy in ways that matter.

The first is speed. Steam power took generations to diffuse. The personal computer needed roughly fifteen years to reshape the office. Generative AI reached a hundred million users within two months of ChatGPT's late-2022 release, and the lag between a capability appearing in a lab and a company deploying it against payroll has compressed from years to months. The World Economic Forum estimates that 39% of the skills workers use today will be outdated or transformed within five years (Future of Jobs Report, 2025). Institutions that once had a generation to adjust now have a business cycle.

The second is breadth—and direction. Every previous wave of automation moved through the economy along a predictable path: it came for muscle and routine first, hands before heads. This one is moving the other way.

The Numbers, and What They Actually Count

Start with the projection everyone cites. The World Economic Forum's Future of Jobs Report 2025, built from surveys of more than 1,000 employers representing over 14 million workers across 55 economies, forecasts that by 2030 some 92 million jobs will be displaced and 170 million created—a net gain of 78 million, against a churn equal to 22% of the roughly 1.2 billion formal jobs in its dataset.

Before we lean on that number, we should be honest about what it is. It is not a measurement. It is a survey of what executives say they plan to do, aggregated and projected forward. The famous "85 million displaced" figure from the WEF's 2020 edition, and this cycle's "92 million," share a methodological spine: they are employer intentions, not counted job losses, and employer intentions are notoriously optimistic about the "created" side of the ledger and vague about timing. A net figure of +78 million tells you almost nothing about the person whose specific job vanishes in 2027 and whose town has no use for the 78 million roles appearing somewhere else. The framing of "jobs displaced versus jobs created" is seductive precisely because it nets out the two things that hurt most: distribution and timing. A balanced ledger can sit atop a great deal of individual ruin.

For something closer to measurement, look at the layoff trackers and the payroll data.

Challenger, Gray & Christmas, which has tracked stated reasons for U.S. job cuts since 2023, found AI cited in 101,743 layoff announcements in just the first half of 2026—already nearly double the 54,836 attributed to AI in all of 2025. AI led all stated reasons for job cuts for four consecutive months into mid-2026, peaking at 38,579 cuts in a single month, the highest monthly total ever recorded. These are cautious numbers, because "we're restructuring around efficiency" rarely gets logged as "AI took the work," so the true figure runs higher than the label admits.

Then there is the cleanest evidence we have, and it points straight at the young. Economists led by Erik Brynjolfsson at Stanford, working with high-frequency payroll records from ADP covering millions of American workers, found that employment for workers aged 22 to 25 in the most AI-exposed occupations fell about 13% relative to their less-exposed peers after late 2022—with the decline reaching roughly 20% for early-career software developers—even after controlling for firm-level shocks (Stanford AI Index, 2026). Crucially, employment for workers over 35 in the very same occupations held steady or rose. The machines are not replacing the job. They are replacing the beginner.

Why the Collar Flipped

For decades the conventional wisdom was that a college degree was a moat. Robots would take the factory and the loading dock; the lawyer, the analyst, and the copywriter were safe behind their credentials. That intuition has aged badly, and the reason is worth spelling out, because it explains almost everything about who is exposed today.

Earlier automation was mechanical. It substituted for physical force and manual repetition, so it flowed toward physical, repetitive work. Today's AI is linguistic and cognitive. Its core competencies are language processing, pattern recognition, and a growing capacity for structured reasoning—which means it substitutes most easily for jobs made of words, patterns, and rules. That describes an enormous amount of white-collar work: drafting, summarizing, coding, screening résumés, answering routine questions, reconciling numbers, writing the first version of nearly anything.

Meanwhile the trades sit on the far side of a wall the current technology cannot climb, made of physical dexterity and unstructured judgment in messy real-world settings. A plumber diagnosing a leak behind a wall, a nurse reading a patient's face, an electrician improvising in a hundred-year-old building—these demand exactly the embodied, situational competence that large language models lack. The result is an inversion of the old fear.

Data-entry work now ranks near the top of exposure lists, with analysts estimating something like 95% of its tasks within reach of current tools; customer service, employing millions in the U.S. alone, sits close behind. The plumber who shows up to fix Marcus's burst pipe is, for now, considerably safer than the junior lawyer drafting the insurance claim that follows.

The economic conditions that flipped the switch arrived quickly and together. The cost of running a capable model has fallen by orders of magnitude since 2022. Enterprise tooling—the connective software that lets a company point AI at its own documents, tickets, and codebase—matured from science project to purchasable product. Integration barriers that once required a data-science team now resolve into an API call and a monthly invoice. When a capability is cheap, packaged, and easy to plug in, adoption stops waiting for permission. That, more than any single breakthrough, is why the deployment curve bent upward when it did.

The Beginner's Trap

The concentration of pain on early-career workers is not a rounding error in the data. It may be the most consequential thing happening.

Entry-level jobs are disappearing not through dramatic firings but through a quieter mechanism: companies simply stop backfilling the bottom rung. Surveys in 2026 found roughly a fifth of firms had already frozen entry-level hiring because of AI, with more expecting to follow, and a striking share anticipating that they would eliminate entry-level hiring entirely within a couple of years. No layoff announcement accompanies a job that is never posted.

The immediate cost is lost income for a generation of graduates. The deeper cost is structural. Entry-level roles were never only about the work—they were apprenticeships in disguise, where people learned how organizations actually function, how to handle a difficult client, how to recover from a mistake with a manager watching. The junior lawyer doing document review was learning to think like a lawyer. The junior analyst building the unglamorous model was learning which assumptions break. If AI absorbs precisely those first-rung tasks, we may be dismantling the training ground that produces senior people—hollowing out the pipeline that supplies the experts, managers, and leaders of the next decade.

There is a genuine paradox here that no one has resolved. If the beginner jobs vanish, where do the experienced workers of 2035 come from? Firms are, in effect, harvesting a stock of senior talent they are no longer replenishing. This is one of the sharpest open questions of the transition, and honesty requires admitting we do not yet know how it resolves.

Transformed, Not Erased—and the Fine Print

Not every affected job disappears. Many are being reorganized around the machine, and this is where the "displacement" frame most badly misleads.

Radiology is the standing example. For years it was held up as the profession AI would obliterate, since image recognition is exactly what the technology does well. What happened instead is that AI became a tireless flagger of anomalies, and radiologists shifted their hours toward complex diagnosis, patient consultation, and the treatment decisions machines cannot own. The role evolved. The same pattern is visible in legal research, financial analysis, and software development: AI takes the routine substrate, humans keep the judgment on top.

But "transformed rather than eliminated" is not automatically good news for the worker inside the transformation, and we should resist the temptation to treat it as a happy ending. Whether absorbing AI into a job improves or degrades it depends on who captures the gains. When AI handles the drudgery and frees a professional for higher-value work, job quality can rise. When AI sets the pace, monitors output, and reduces a skilled role to reviewing and correcting a machine's drafts all day, the same "transformation" can hollow out the craft, intensify the workload, and strip away the satisfaction that made the job worth doing. Early evidence cuts both ways. The evolution of a role is a distributional question wearing a technical costume—it depends entirely on whether the worker is elevated by the tool or chained to it.

This nuance also explains a puzzle. Despite three years of acute anxiety since generative AI went mainstream, the aggregate labor market has not collapsed. Unemployment in most advanced economies has stayed historically low, and PwC's 2026 analysis found that headcount at the most AI-exposed companies actually grew faster than at the least-exposed ones. The disruption is real, accelerating, and brutally uneven—but so far it has been a redistribution of who works and at what level, not a wholesale disappearance of work.

The New Jobs, and the Catch

AI is unmistakably creating work. Job categories that barely existed five years ago—machine-learning engineers, AI product managers, model-evaluation specialists, AI ethicists and governance leads, applied-AI architects—are among the fastest-growing on the market. PwC's 2026 Global AI Jobs Barometer found that jobs demanding AI skills are growing several times faster than the market overall, and the wage premium for those skills has climbed to 62%, up from 57% a year earlier. Average pay for AI engineers reached roughly $206,000 in 2026. This is a genuine boom, not a mirage.

The catch is not that the jobs are fake. The catch is that they land on different people, in different places, at different rungs than the jobs being lost.

Consider what happened to "prompt engineer," the poster-child new job of 2023. By 2026 the standalone title was appearing less often on job boards than it had two years earlier—not because the work vanished, but because it dissolved into the expected skillset of AI engineers, applied ML engineers, and solutions architects. New job categories are unstable in exactly this way, which makes the "170 million new roles" projection genuinely hard to plan a career around. We do not reliably know what many of these roles will require in five years, because the roles themselves keep mutating.

The credential picture is contested, and worth being careful about. Earlier datasets suggested the new AI roles were locked behind graduate degrees—one widely cited figure held that the overwhelming majority required a master's or doctorate. More recent industry hiring data complicates that: nearly half of AI-engineering postings in 2026 accepted a bachelor's or master's and emphasized demonstrable, practical experience over formal pedigree. The honest synthesis is that the ceiling has cracked but the floor has risen. You may not need a PhD, but you almost certainly need substantial technical fluency—and that is still a chasm, not a step, for a displaced customer-service representative with a high-school diploma or a junior accountant whose degree is in accounting. The premium is real. Access to it is not evenly distributed.

Which produces the defining feature of the emerging labor market: a fork. PwC's data describes two diverging tracks—jobs "professionalised" by AI, where the technology amplifies a worker's expertise and wages grow rapidly, and jobs "democratised" by AI, where the technology commoditizes a task and wage growth stalls. Since 2021, the professionalised track has seen wages grow markedly faster. The same technology is simultaneously a ladder for some and a floor lowered onto others, and which one you experience depends heavily on where you started.

Communities, Not Just Careers

Displacement does not fall evenly across the map, and its second-order effects ripple outward from the individual to the town.

When earlier automation gutted manufacturing regions, the damage was never confined to the laid-off worker. It moved through the local economy—the diner, the barber, the school district funded by a shrinking tax base—and it settled into a durable regional decline that outlasted the original shock by decades. AI's exposure map is different from manufacturing's, which changes which places are vulnerable. Because this wave targets cognitive and administrative work, the exposed geography includes back-office hubs, customer-service centers, and the mid-sized cities whose economies rest on white-collar processing work rather than factories. A metro area built around insurance-claims processing or call centers now faces the kind of concentrated shock that once hit steel towns—and it faces it without the folk memory or the policy playbook that manufacturing decline eventually, painfully, generated.

What History Can and Cannot Promise

The long view offers real reassurance and a real warning, and we should take both.

The reassurance: across two centuries, technological revolutions created more total jobs and more total wealth than they destroyed. The warning: the transitions took decades, and the people living through them absorbed genuine suffering that "in the long run it worked out" does nothing to redeem. The weaver did not get to live in the long run. He lived in 1815.

The distinctive risk this time is not permanent, economy-wide unemployment—the balance of evidence still points toward net job creation. The distinctive risk is that the pace of change outruns the speed at which people and institutions can adapt. When the compression of time meets the breadth of exposure meets the concentration on the young, the adjustment burden falls hardest on individuals who have the least capacity to bear it, and it falls before the institutions meant to cushion it have even understood the shape of the problem.

Over the next five to ten years, the most exposed categories—routine administrative, entry-level analytical, and high-volume customer-facing roles—will likely keep contracting under any plausible scenario. Under a baseline of steady adoption, that contraction is gradual enough for retraining and attrition to blunt some of the edge. Under an accelerated scenario, where capable AI agents move from assisting to autonomously executing multi-step work, the contraction could arrive faster than any existing safety net is built to absorb. The difference between those two futures is not mainly technological. It is a matter of choices we have not yet made.

The Question We Actually Have to Answer

Which brings us to the part no dataset can settle. The productivity gains are real and large—PwC found the most AI-exposed sectors posting far higher productivity growth than the least-exposed ones. The question is who receives them. When a machine does the work a person used to be paid for, the value that person created does not evaporate; it moves. By default, it moves to whoever owns the machine. Whether it is shared with the displaced worker—through wages, retraining, redistributed gains, or new work—is not decided by the technology. It is decided by policy, bargaining power, and political choice. This is the distribution question, and in this domain it is the question.

The policy toolkit on offer is familiar and each instrument comes with a real trade-off. Retraining subsidies are the default answer, but the evidence on large-scale retraining is sobering: programs are often short, generic, and disconnected from what employers actually need, and asking a fifty-year-old claims processor to become an AI-fluent analyst in a twelve-week course is a promise the data rarely keeps. Expanded unemployment insurance and portable benefits—support that follows the worker rather than the job—buy time and dignity during transition but do not by themselves create the next job. Deeper education reform, aimed at building adaptable skills earlier, may be the most durable lever, but it pays off over a generation, which is precisely the time this transition is not giving us. There is no clean answer, only a set of imperfect instruments whose value depends on deploying them before the disruption crests rather than after.

And beneath the toolkit sits an unresolved question of responsibility. Who owes the displaced worker a future—the government whose tax base the productivity gains inflate, the employer who captured the savings, the technology developer who built the substitute, or the individual expected to reskill on their own initiative? Each actor has a plausible claim to duck, and in the ducking lies the danger. The transitions of the past turned brutal not because anyone chose cruelty but because everyone assumed adaptation was someone else's job. The Luddites lost not because they were wrong about the machines, but because no one with power had decided the displaced were their responsibility.

That, in the end, is the choice in front of us. The technology will keep improving on its own schedule. Whether its gains widen the intelligence divide or narrow it is not a fact we will discover. It is a decision we will make—or, if we look away, a decision that will be made for us by default.

Summary

1. The pattern is old; the pace is not. Machines have displaced workers before—the Spinning Jenny, the railroads, computerized clerical work—and the economy eventually created more jobs than it destroyed. What breaks the analogy today is speed and breadth: generative AI diffused in months, not decades, and the WEF estimates 39% of current skills will be transformed within five years.

2. Displacement is real and measurable, and accelerating. AI was cited in over 100,000 U.S. job cuts in the first half of 2026 alone, already nearly double the previous year's full total (Challenger, 2026)—and stated reasons undercount the true figure.

3. The collar flipped. Because today's AI substitutes for language, pattern, and rule-based work rather than physical labor, white-collar and knowledge roles are proving more immediately exposed than the trades. The plumber is safer than the paralegal.

4. The young bear the burden. Employment for workers aged 22–25 in exposed occupations fell about 13% relative to older peers after 2022, while older workers in the same jobs held steady (Stanford, 2026). The disappearing entry-level rung threatens the entire pipeline that produces future experts and leaders.

5. Transformation is not automatically good. Many roles evolve rather than vanish as AI absorbs routine tasks, but whether that improves or degrades a job depends on whether the worker is elevated by the tool or chained to it—a distributional question, not a technical one.

6. The new jobs are real but land elsewhere. AI-skill roles are growing fast and pay a 62% wage premium (PwC, 2026), but they arrive at different rungs, in different places, for different people than the jobs lost—splitting the market into a "professionalised" track that rises and a "democratised" track that stalls.

7. The headline projections are intentions, not measurements. The WEF's "92 million displaced, 170 million created" figures come from employer surveys and net out the two things that hurt most: distribution and timing. A balanced ledger can sit atop widespread individual ruin.

8. The distribution question is the question. The productivity gains are large and real; who receives them is undecided by the technology and left to policy, bargaining power, and choice. Retraining, portable benefits, and education reform each help and each fall short—and the transition turns cruel mainly when everyone assumes the displaced are someone else's responsibility.

Sources

• 77 AI Job Replacement Statistics 2026 (New Data)
• AI Job Displacement 2025: Which Jobs Are At Risk?
• Top 20 Predictions from Experts on AI Job Loss
• 59 AI Job Statistics: Future of U.S. Jobs | National University
• How will Artificial Intelligence Affect Jobs 2026-2030 | Nexford University
• AI Job Loss Statistics 2026: Who's Losing, Who's Hiring, etc. | SQ Magazine
• I Analyzed 76,000 AI Job Losses in 2025
• SHRM Research: AI Impact on HR Employment
• AI's Wake-Up Call: 23.2 Million American Jobs Already Impacted
• Why AI Will Create More Jobs Than It Will Eliminate
• Over 97 Million Jobs Set to be Created by AI
• AI Is Making Jobs, Not Taking Them | RAND
• New Skills and AI Are Reshaping the Future of Work | IMF
• Visualizing the Top 40 Jobs at Risk From AI
• Microsoft Research: 40 Jobs Most Exposed to AI | Fortune
• The Dawn of Automation: A Historical Perspective
• A Short History of Jobs and Automation | World Economic Forum
• Machinery and Labor in the Early Industrial Revolution | MIT

Last updated: 2026-07-02