Mass Unemployment

Picture Daniel Foster in the autumn of 2033. He is forty-two, holds a master's degree in finance, and has not held steady work in three years. This is not a story about a man who stopped trying. He has sent out more than four hundred applications since 2030 and sat for twenty-three interviews. He has zero offers. The reason is not Daniel. The reason is that the job he spent fifteen years mastering — corporate accounting — has been substantially handed to software.

His old department told the story in three acts. In 2028 his firm installed an AI ledger-and-audit system and cut the accounting team from twenty-four people to eight. Daniel survived. In 2029 a further upgrade took it to three. Daniel did not. "Your skills are valuable," his manager said, in the peculiar consoling register that managers had learned to use by then, "but the system does what you did, and we can't justify the headcount." Every other firm was installing the same systems. The rare openings asked for "AI oversight experience" — the skill of supervising the machine that had replaced him — and went to twenty-six-year-olds. By 2031 Daniel was packing boxes in a warehouse for eighteen dollars an hour, a third of his old salary, until the warehouse automated too. Delivery driver, chatbot-displaced customer service, data entry gone to a model. In 2033 he is a member of what a handful of economists have started calling the AI precariat.

Daniel is a composite, and this chapter is openly a scenario — a disciplined attempt to follow present-day forecasts to one plausible 2033 and ask what it would feel like from inside. The value of the exercise is not the specific man but a specific gap: the distance between what the careful institutions predicted in the mid-2020s and what this scenario suggests actually arrives. That gap turns out to be the most important thing in the story, because it runs in one direction. Almost every early forecast, it seems, was too low.

The forecasts, and how they aged

Go back to the mid-2020s and read the credible projections. Forrester estimated that automation and AI would eliminate around 6 percent of US jobs by 2030 — roughly 10 million positions (Forrester, 2025). The World Economic Forum and allied bodies put global displacement somewhere in the range of 85 to 92 million jobs by the end of the decade, typically paired with a reassuring second clause about tens of millions of new jobs created alongside them (WEF, Future of Jobs). Goldman Sachs framed it as exposure rather than loss, estimating that generative AI could expose the equivalent of 300 million full-time jobs worldwide to some degree of automation (Goldman Sachs, 2023). These were the sober numbers, the ones that made it into policy briefings.

In this scenario they proved conservative. By 2033 the US figure looks closer to 12 percent than to 6, and the global count of eliminated positions passes 120 million rather than resting at 90. The reassuring second clause — the new jobs — did not arrive at the promised volume. What the early models missed was not the direction but the coupling: they assumed AI capability would improve at one pace and diffuse into firms at another, slower one, leaving a comfortable lag in which displaced workers could reabsorb. Instead capability and deployment advanced together, and the lag collapsed.

The pattern of the miss is itself instructive. Forecasts that hedged toward the low end were not being careless; they were doing what forecasts responsibly do, anchoring to historical rates of technological diffusion. When a projection built on the last two centuries of automation comes in low by a factor of two, the honest conclusion is not that the forecasters were foolish but that the reference class may have been wrong — that this transition does not belong in the same file as the ones before it. Hold that thought; the whole chapter turns on it.

Where the losses concentrated

Displacement did not fall evenly. It arrived in some sectors as a cliff and in others as a slow erosion, and the difference tracks a simple property: how much of the work was language, numbers, and rules — the native territory of large models — versus physical presence and unstructured judgment.

Two things about this table matter more than any single row. The first is that it describes concentrated, not gradual, loss in its top half: customer service did not shed 5 percent a year for sixteen years, it fell off a ledge over roughly thirty months once voice models crossed a quality threshold. The second is that the table is a map of where the primary automation events happened — and the primary events are not where most of the pain ends up.

That is because displacement cascades. When a mid-sized financial firm automates its accounting floor, it does not only shed accountants. It stops leasing the floor, so a commercial landlord loses a tenant. It cancels the cleaning contract, the catering account, the office-supply order, the parking spaces. The sandwich shop downstairs loses its lunch rush; the accountants who ate there are home now, and not eating out. Each of those losses lands in a sector the original automation never touched — real estate, food service, local retail — and lands on workers who never met an AI system. Trace the unemployment far enough downstream and it becomes unrecognizable as an AI story at all. It looks like a struggling neighborhood, which is exactly how it is experienced.

A fair question is where numbers like "~80% customer service" even come from, and the honest answer is that they are engineered estimates carrying real uncertainty, not measurements. Three methods sit behind them. The first is occupational task analysis: labor economists — following the approach Frey and Osborne pioneered in 2013 and the O*NET-based methods that refined it — decompose a job into constituent tasks and ask what share a current model can perform. The second is capability benchmarking: tracking how AI systems score on standardized tests of the underlying skill and projecting the curve forward. The third is employer surveys, simply asking firms what they have automated and intend to. Each has a characteristic bias. Task analysis tends to overstate displacement, because a job is more than the sum of its listed tasks — the tacit coordination between them is often the hard part. Surveys tend to understate it, because firms are cautious about announcing layoffs. And every one of these methods is acutely sensitive to a single assumption: the AI capability trajectory. Shift the assumed rate of improvement, and an "~80%" becomes a "~50%" or a "~95%." The table is best read not as a verdict but as the midpoint of a wide and assumption-dependent band.

Why the machine breaks the old cycle

The most disorienting feature of this scenario is that it coexists with apparent prosperity. GDP rises. Corporate profits are strong. Measured productivity hits records. By the dashboard that governments watch, the economy looks healthy — and employment falls anyway, not as a passing dip but as a lasting condition.

This is a structural decoupling, and to see why it matters you have to see what it severs. For most of the modern era, growth and employment moved together in a self-feeding loop: firms grew and hired, workers earned and spent, that spending drove more growth, and the growth demanded more hiring. The entire architecture of the twentieth-century economy — Keynesian demand management, the postwar welfare state, collective bargaining, payroll-funded social insurance — was built on the assumption that this loop would hold, that prosperity would pass through labor on its way to consumption. AI lets a company step off the loop. It can now expand output by adding compute instead of employees, capturing the productivity gain without ever routing it through a wage. Labor stops being the thing that scales with demand and becomes one input among several — and often the most expendable one.

Follow the money and the danger sharpens into a feedback risk. The workers automation displaces are also customers. Their wages were demand — for restaurants, retail, haircuts, gym memberships, the whole texture of local commerce. Pull those wages out and consumption contracts even as corporate revenue climbs, which pressures the surviving businesses to cut costs, which they increasingly do by automating, which removes more wages. Whether this tips into a self-reinforcing contraction depends on a threshold economists cannot locate precisely in advance: it depends on how concentrated the losses are, how much of displaced workers' spending is replaced by transfers, and how quickly. An economy with strong automatic stabilizers and generous unemployment support sits far from the edge. One where displacement outruns the safety net moves toward it — and the unnerving property of the loop is that you generally confirm you have crossed the threshold only after you have.

The retraining answer, and why it isn't one

Every prior wave of displacement summoned the same institutional reflex: retrain the workers for the jobs the new technology is creating. The reflex was rational, because mechanization, electrification, and computerization each did, eventually, create more jobs than they destroyed. Retraining became the load-bearing assumption of labor policy across the developed world. In this scenario it fails, and it fails for three independent reasons, any one of which would be sufficient on its own.

The first is pace. Retraining has always raced the clock between losing an old job and being able to compete for a new one. In earlier transitions the destination skill stayed valuable long enough to repay the investment. Here the window keeps slamming shut. Workers who enrolled in 2030 to learn AI-assisted bookkeeping found in 2032 that AI-assisted bookkeeping was itself being automated. You cannot train your way onto a moving target when the target moves faster than the course catalog.

The second is scale, and it is arithmetic, not effort. Even flawless individual retraining cannot place tens of millions of people into roles that do not exist in that number. The genuinely AI-resistant technical fields — model development, robotics maintenance, systems oversight — employ a comparatively tiny workforce. The new jobs are real; there are simply nowhere near enough of them. Telling ten million displaced clerks to become machine-learning engineers is like telling the passengers of a sinking ferry to become lifeboat builders: true in principle, useless at the relevant ratio.

The third is a cluster of structural barriers that survive even successful training. The new roles concentrate in a handful of expensive cities while displaced workers are spread across regions that cannot all pick up and move. Age discrimination persists past the certificate, as fast-moving technical employers quietly prefer candidates who grew up with the tools. And the cognitive demands of AI-adjacent work — sustained abstract reasoning, comfort inside constantly shifting systems — are not uniformly trainable on the timeline that a lost mortgage payment imposes. Retraining still helps particular people, and it belongs in any serious policy mix. But as a system-level answer to displacement at this scale, it has functioned more as a way to look responsive than as a solution.

The fracture

Statistics carry the scale of this but not its weight, and the weight falls along several fault lines that reinforce one another.

Geography goes first and most visibly. Regions built around the automatable middle — back-office finance, routine legal processing, clerical and administrative work — contract faster than any local institution can adjust. The workers who can afford to leave do, taking the remaining skilled labor and the tax base with them, which deepens the hole for everyone who stays. Property values fall, schools close, municipal budgets strain, services thin out, which pushes more people to leave: the ordinary vicious spiral of a declining town, now running in places that were comfortable a decade earlier.

Generations fracture in opposite directions at once. Older workers lose not just income but the professional self built over decades — an identity demolition that no severance check addresses. Younger workers, meanwhile, cannot start the careers they trained for, because the entry-level rungs have been sawn off; graduate unemployment in the hardest-hit fields runs far above historical norms in this scenario. And the cohort in the middle, in their forties and fifties, is arguably worst placed of all: too experienced to accept entry wages, too specialized in now-automated skills to pivot cheaply, too young to retire. Displacement hits both ends of a career simultaneously, which is historically unusual and politically combustible.

Radicalization tends to follow economic desperation with grim reliability, and this scenario is no exception. Movements offering a clear villain — immigrants, elites, the tech companies, globalization — gain among people who kept their side of the social bargain and got unemployment in return. Their grievance is legitimate even where the politics it feeds are not. And underneath all of it runs the psychological layer: in a society that organizes status, structure, and self-worth around a job, structural unemployment removes the scaffolding of a life, not merely its income. Depression, shame, and withdrawal spread through the long-term unemployed, straining exactly the mental-health and social services that shrinking local budgets can least afford.

Here the fracture rejoins the paradox from earlier, and does its most specific political damage. When leaders stand up and celebrate record GDP and record profits to constituents living through what feels like a depression, the constituents do not conclude that they have misread the numbers. They conclude that the numbers are a lie told by people who do not see them — and, crucially, they are not wrong. The headline statistics and the individual experience are telling genuinely different true stories, because the aggregate has stopped being representative of the median. Averages survive when the top rises fast enough to cover a falling middle. The instrument still reads "healthy," but it is measuring a body most people no longer live in, and every celebration of the reading spends down the last reserves of institutional trust.

The work that stays human

Not everything automates, and the pattern of what survives is legible. Durable roles cluster around four properties.

High-skill creative and strategic work — research scientists, architects, senior strategists, executives — persists where value depends on originality, judgment under genuine novelty, and the integration of knowledge across domains the model was never handed cleanly. Skilled trades — electricians, plumbers, HVAC technicians, carpenters — persist because the physical world they work in is too variable and unstructured for general-purpose robotics to handle affordably at scale; it remains cheaper to send a human into a cramped, non-standard crawlspace than to build a robot that can. Interpersonal care — in-person teaching, nursing, therapy, childcare — persists partly for technical reasons and partly because people keep a real preference for human presence in moments of vulnerability. And legally accountable roles — judges, regulators, compliance officers — persist because the law demands a human answerable for consequential decisions, and because the legitimacy of those decisions rests on human authorship.

These are genuine categories doing meaningful work. They are also nowhere near large enough. There is no world with a hundred million openings for master electricians and ICU nurses, and no amount of aggressive training conjures those positions into existence quickly. What survives is, on average, harder to enter, more geographically concentrated, and less evenly accessible than what was lost — which means the surviving work cannot function as the reabsorption engine that every previous transition eventually provided.

Is this time actually different?

Everything above rests on one contested claim, so it deserves the strongest version of the counterargument. The optimist's case is not naive; it is empirical. Mechanization, electrification, and computerization each destroyed whole categories of work and each ended with more employment than before, because each freed human effort for something the machine could not do. The tractor ended most farm labor and the displaced went into factories; the factory automated and the displaced went into offices and services. In every case the machine took over a physical or narrowly routine task and, in doing so, raised the value of the human capacities it couldn't touch — judgment, coordination, care, creativity. Automation kept elevating the same reservoir of distinctly human ability, and the economy kept expanding into it.

The reason this time may genuinely differ is that AI competes inside that reservoir. For the first time the technology's frontier is cognitive: it is coming for judgment, analysis, language, and increasingly the creative and coordinating work that every previous wave pushed people toward. When the tractor arrived, the exit door was cognitive work, and it was wide open. If AI narrows that door while automating the physical one through robotics, it is not obvious what the next room is. The pattern of prior transitions depended on a specific mechanism — machines complementing human cognition by handling everything else — and it is precisely that mechanism AI puts in question.

This is the pivot on which honesty is required, and where the book's registers matter most. That new categories of human contribution have always emerged before is established history. That they will emerge again is a hope, not a finding — and the specific mechanism that produced them last time is the one thing AI most directly threatens. Anyone who tells you confidently that the jobs will return, or confidently that they won't, is claiming knowledge that does not exist. The intellectually honest position is that we do not know, and that the not-knowing is not a rhetorical dodge but the actual state of the evidence.

Policy under deep uncertainty

If the central fact is uncertainty, then the first test of any policy is not whether it fixes the optimistic world or the pessimistic one, but whether it survives being wrong about which world we are in. Measured against the underlying problem — that market economies distribute income mainly by economic contribution, and AI is systematically stripping people of contributions to be paid for — the leading responses each address a different slice, with different obstacles.

Job guarantees offer public employment to anyone displaced, delivering structure and income. But the roles are easily perceived as make-work, and sustaining tens of millions of them without crowding out private investment demands either heavy borrowing or steep taxes, both politically fraught. Robot taxes — levying firms that swap workers for AI and routing the revenue to the displaced — are structurally undercut by competition: a company taxed for automating in one country can move to one that doesn't tax it, so no single nation can push the rate far without international coordination that has not materialized. Work-sharing mandates spread shrinking hours across more people, keeping more of them attached to jobs, but they leave the income gap from those lost hours unaddressed and run against the very productivity logic that drives automation in the first place.

Universal basic income is the one response that attacks the core mismatch directly, because it severs income from contribution by design. And here the evidence deserves precision rather than enthusiasm, because it is thinner than advocates imply. The real pilots were small and short. Finland ran roughly 2,000 unemployed people on about €560 a month for two years and found improved wellbeing but no clear employment effect. Stockton's SEED program gave 125 people $500 a month for two years and saw full-time employment rise among recipients. GiveDirectly's Kenyan programs, larger and longer, showed durable gains in wellbeing and local economic activity. The most rigorous — the OpenResearch study backed by Sam Altman — gave 1,000 low-income Americans $1,000 a month for three years and found modestly reduced work hours alongside real improvements in financial security and autonomy.

The trouble is that none of these can answer the question a national program actually raises. A pilot hands money to a few thousand people inside an otherwise normal economy; every recipient can still spend their transfer in a functioning labor market surrounded by employed neighbors. Scale that to a whole country and you change the system the pilot was measured against — you alter prices, labor supply, inflation expectations, and the tax base funding it all, none of which a local experiment can observe. Short-term, small-scale pilots are genuinely encouraging about the effect on individuals and genuinely silent about the macroeconomics of nationwide implementation. Treating the first as evidence for the second is the single most common error in the UBI debate.

Which leaves the hardest normative question: who should pay for a structural shift no individual chose? Three answers each have a real argument. Put the cost on the firms that capture the productivity gains, on the polluter-pays logic that whoever profits from imposing a transition cost on society should internalize it. Put it on the public broadly, on the grounds that automation's cheaper goods and services benefit everyone as consumers, so everyone should fund the adjustment. Or — the position almost no one defends openly but the default we drift into by inaction — leave it on the displaced workers and their communities, on the market's blunt claim that outcomes reflect contribution. The strongest case belongs to the first: the gains are concentrated and identifiable, the costs are currently socialized onto people who did nothing wrong, and leaving that asymmetry in place is the choice most corrosive to the belief that the system is fair.

Given that we cannot know in advance whether the new jobs return, the design principle that follows is robustness rather than optimization. Build institutions that are tolerable if the optimists are right and survivable if the pessimists are — a decoupling of basic security from employment that, in the good scenario, costs a manageable sum for a safety net that mostly goes unused, and in the bad scenario, is already standing when it is needed most. Funding tied to where the productivity gains actually accrue, benefits that don't assume a return to full employment but don't punish work if it returns, and enough international coordination to keep the mobile tax base from dissolving the whole arrangement. The point is not to bet on a forecast. It is to stop needing the forecast to be right.

The outlook, and the honest ledger

The near-term official numbers understate all of this, almost by construction. Standard projections show US unemployment running perhaps a million higher in 2035 than in 2030, peaking around 1.5 million additional unemployed near 2040 — already unusual, and already conservative if the capability curve of the late 2020s holds. Extend that curve and unemployment in the 20 to 30 percent range across advanced economies by the mid-2030s stops being alarmist and becomes one plausible branch. Add the consumption feedback loop, and the branch bends toward something a labor market has not seen since the 1930s.

None of these figures is a prediction in the sense of a promise. They are the visible arithmetic of assumptions that may or may not hold, and the largest of those assumptions is the one this chapter has circled throughout: whether the historical pattern of new work creation repeats. What we know is that the twentieth century's institutions — its safety nets, its payroll-funded insurance, its retraining reflex — were built for a coupled economy that AI is decoupling. What we can reasonably expect is that displacement will keep running ahead of the tools built to absorb it for as long as capability and deployment advance together. And what we can only hope, or fear, is the rest.

Summary

1. The early forecasts were conservative, and consistently in one direction. Where credible mid-2020s projections put global displacement at 85–92 million jobs and US losses near 6% by 2030, this scenario reaches roughly 120 million and about 12% — a miss by nearly a factor of two, driven by capability and deployment advancing together instead of with the assumed lag.

2. The crisis is structural, not cyclical. These jobs do not return when conditions improve, because the work has been permanently reassigned to machines.

3. Growth and employment have decoupled. Firms expand output by adding compute rather than workers, breaking the wage-to-consumption loop that twentieth-century institutions were built to sustain — and raising a real risk of a self-reinforcing contraction as lost wages cut consumer demand.

4. Retraining fails for three independent reasons: skills obsolesce faster than courses can teach them (pace), there are far too few destination roles (scale), and geography, age, and cognitive demands block even successful graduates (structural barriers).

5. The damage cascades and compounds. Primary automation events trigger secondary losses in real estate, retail, and local services, and the social fractures — geographic collapse, generational squeeze, radicalization, psychological harm — reinforce one another.

6. The statistics and the experience diverge honestly. When averages stay healthy because gains at the top mask a falling middle, celebrating the headline numbers is not spin — it is measuring a body most people no longer live in, and it spends down institutional trust.

7. The work that stays human is real but too small. Creative-strategic, skilled-trade, interpersonal-care, and legally-accountable roles persist for identifiable reasons, but cannot absorb displacement at the required scale.

8. The sector estimates are engineered, not measured. Task analysis, capability benchmarking, and employer surveys each carry known biases and are highly sensitive to assumptions about the AI capability trajectory; read the numbers as wide bands, not verdicts.

9. The UBI evidence is encouraging about individuals and silent about the macroeconomy. The pilots were small and short (Finland, Stockton, GiveDirectly, OpenResearch); nationwide implementation changes the very system the pilots were measured against.

10. "This time is different" is the central open question, and it is genuinely open. Prior automation elevated human cognition; AI competes in it, threatening the exact mechanism that created new work before. The right response is to design institutions robust across both outcomes rather than to bet on either — and, where costs must fall, to place them on those capturing the gains rather than on the communities that never chose the transition.

Sources

• AI and Automation Will Take 6% Of US Jobs By 2030 | Forrester
• How Will AI Affect the Global Workforce? | Goldman Sachs
• The Impact of AI on the Labour Market | Institute for Global Change
• Evaluating the Impact of AI on the Labor Market | The Budget Lab at Yale
• AI exposure predicts unemployment risk | PMC
• AI and technological unemployment | ScienceDirect
• Top Predictions from Experts on AI Job Loss | AIMultiple
• AI Expert Warns 99% Will Lose Jobs by 2030 | Final Round AI

Last updated: 2026-09-10