The Question Society Keeps Avoiding

What happens when machines can do most jobs?

Not all jobs, and not immediately. But enough jobs, fast enough, that the question stops being hypothetical.

Every domain chapter in this book describes the same underlying move. Work that required human cognition becomes automatable. Output per worker rises. Fewer people are needed to produce the same result. Considered purely as economics, this is good news—it is the definition of productivity growth, and productivity growth is what made modern prosperity possible.

But societies are not organized around production. They are organized around employment. Income reaches households through wages. Days are structured by work schedules. Health insurance, in the United States, is attached to jobs. Status, identity, and much of adult social life derive from occupation—which is why "what do you do?" is the second question strangers ask each other.

Work is the distribution mechanism, the scheduling mechanism, and the meaning mechanism, all at once. Automating it does not simply change the economy. It removes the thing three other systems were quietly resting on.

2026 Snapshot — Work and Automation

The Labor Market as It Stands

Roughly 3.5 billion people are employed worldwide, with hundreds of millions more seeking work or underemployed. US employment sits near 160 million with unemployment around 4 percent—a figure that has stayed low enough, long enough, that it has become an argument in itself: whatever automation has done so far, it has not produced mass joblessness.

That argument is correct and incomplete. Median real wages in developed countries have been close to flat for decades even as output per hour has risen substantially. The gains went somewhere; they did not go to the median worker. And headline unemployment obscures a slower structural change: prime-age male labor force participation in the US has fallen from about 97 percent in 1954 to roughly 89 percent today.⁵ Those men are not counted as unemployed. They are counted as not looking.

That decline began long before modern AI and has multiple causes. It matters here because it establishes that a labor market can shed people quietly, over decades, without ever producing the crisis headline that would force a policy response.

What Automation Has Already Done

Manufacturing automation is largely complete for repetitive tasks, and US manufacturing employment has been declining for forty years while manufacturing output has risen. Service-sector automation—retail checkout, food service, call centers—is well advanced. Knowledge work is where AI is now arriving, and so far the dominant pattern is assistance rather than replacement.

The most-cited estimate comes from McKinsey, which puts around 30 percent of hours currently worked in the US as automatable by 2030 with technology that already exists.¹ That figure is frequently misreported as 30 percent job losses, which it is not. Hours are not jobs. A role that becomes 30 percent automatable usually becomes a role with different tasks, not a vacancy.

Whether that stays true is the entire question.

Task Versus Job

The distinction that governs everything in this chapter: most jobs contain automatable tasks, and very few jobs are entirely automatable.

A radiologist reads images, consults with clinicians, handles ambiguous cases, manages a department, and reassures patients. AI is superhuman at exactly one of those. A paralegal reviews documents, drafts filings, manages deadlines, and talks to clients. AI does the first two well.

This is why the near-term effect is compositional rather than binary. Jobs do not vanish; they lose their routine content and retain their non-routine content. The consequences follow from that: fewer people are needed per unit of output, entry-level positions—which are disproportionately routine—thin out first, and the ladder that junior workers climb to become senior workers loses its bottom rungs.

That last effect is underappreciated. An industry can look healthy in aggregate while quietly ceasing to produce the next generation of its own experts.

Notable Perspectives

The Pessimists

Technological unemployment is an old worry. Keynes named it in 1930 and predicted it would be a temporary phase of maladjustment on the way to a fifteen-hour work week. He was right about the productivity and wrong about the hours.

The serious modern version does not claim that this time nothing will be left. It claims that this time the substitution targets cognition itself, which is the faculty humans have always retreated to when machines took over physical work. When the loom took weaving, weavers moved to work requiring judgment. If judgment is also automated, the question of where displaced workers move has no obvious answer—not because no answer exists, but because the historical pattern that generated previous answers no longer applies.

The pace argument is separate and stronger. Agricultural employment in the US fell from about 40 percent of the workforce to 2 percent over the twentieth century, and manufacturing from 25 percent to under 10 percent.⁹ Both transitions were absorbed. Both took generations, which meant that most of the adjustment happened through people entering different work rather than leaving their own. A transition compressed into fifteen years does not offer that mechanism.

The Optimists

Every previous automation wave destroyed occupations and created more than it destroyed, and the created jobs were mostly unimaginable beforehand. Nobody in 1990 forecast a labor market containing millions of people doing search engine optimization, cloud infrastructure, or content moderation. The inability to name future jobs is not evidence that there will be none; it is the expected condition.

The complementarity argument is the strongest optimistic case. Tools that raise a worker's output raise the value of that worker's remaining judgment. If AI handles the routine 60 percent of a job, the human's contribution to the other 40 percent becomes more valuable per hour, not less.

The Realists

Both stories can be true at once, and usually are. Aggregate long-run outcomes have historically been good; the transition costs have historically been borne by specific people who did not recover. The English handloom weavers were right about their own lives even though the optimists were right about the century.

Distribution is the crux. Aggregate gains do not comfort someone whose occupation disappeared at fifty-two. And whether the transition is cushioned or not is a policy choice, not a property of the technology.

The Displacement Trajectory

Automating first: routine cognitive work with clear inputs and verifiable outputs. Data entry, first-pass document review, standard financial analysis, template-driven writing, scheduling, tier-one customer support, basic code generation, transcription and translation.

Automating later: work requiring judgment under genuine uncertainty, physical dexterity in unstructured environments, and complex interpersonal negotiation. A plumber diagnosing a problem in an unfamiliar hundred-year-old house is doing something considerably harder for a machine than most white-collar work.

Possibly never, for reasons that are not technical: roles where people want a human specifically—childcare, eldercare, therapy, hospitality, live performance, competitive sport. Roles that are legally reserved. Roles where accountability requires a person who can be held responsible, which is a legal and moral requirement rather than a capability gap.

That last category is larger and more durable than technologists tend to assume. The demand for a human in the loop is not always a demand for competence. Sometimes it is a demand for someone to answer to.

Policy Responses

Universal Basic Income

An unconditional periodic cash payment to every citizen regardless of work or means.

The case for it is that it decouples survival from employment, which is precisely the coupling that automation threatens; that it eliminates the administrative apparatus and the humiliation of means-testing; and that it does not create the poverty traps that benefit phase-outs produce.

The evidence base is broader than critics usually acknowledge and weaker than advocates usually claim. Finland's 2017–2018 trial paid €560 a month to 2,000 unemployed people and found reduced stress and improved wellbeing, no reduction in job-seeking, and a small employment increase.³ GiveDirectly's Kenya program—the largest and longest-running, paying roughly $22 a month to more than 20,000 people over twelve years—shows durable welfare improvements and, importantly, no evidence of the idleness effect that dominates political objections.⁴ Stockton's pilot showed increased full-time employment among recipients.

The critical caveat is that none of these tested UBI. They tested cash transfers to selected groups, funded externally, at small scale, for limited periods. A genuine universal program would be funded by taxation, would be permanent, and would change prices and wage expectations economy-wide. Those macro effects are precisely what the pilots cannot measure, and they are where the real disagreement lies.

The cost objection is arithmetic and unavoidable. A meaningful US UBI runs into the trillions annually—comparable to the entire existing federal budget. It is not impossible to finance, but it requires a tax structure no democracy has yet enacted absent total war.

Negative income tax, Milton Friedman's proposal, addresses the cost by phasing benefits out as earnings rise, delivering a floor at a fraction of universal-payment cost.⁷ It reintroduces means-testing and marginal-rate distortions, and it is the design most likely to be politically achievable precisely because it resembles what already exists.

Job Guarantee

Rather than paying people not to work, the state acts as employer of last resort. India's MGNREGA guarantees 100 days of wage employment per rural household per year and employs more than 50 million people annually—the largest such program ever run, and evidence that the design is administratively feasible at scale.⁸ New Deal programs are the historical precedent.

The advantage is that it preserves work's structure and social role while meeting real public needs—care, maintenance, environmental restoration—that markets underfund. The difficulty is that guaranteeing a job means specifying the job, and a program whose work is visibly unnecessary undermines the dignity it was designed to provide.

Wage Subsidies

The Earned Income Tax Credit is the largest US anti-poverty program for working households, providing up to roughly $7,000 a year to qualifying families.⁶ Its record on increasing employment and reducing child poverty is strong.

Its limitation in this context is structural: it makes work pay more, which requires that work exist. As a response to automation it addresses the wrong variable.

Robot Taxes

Taxing automation to slow substitution and fund transition. The intuition is reasonable—the tax code currently favors capital investment over labor, which subsidizes exactly the substitution in question. The implementation is the problem, since "robot" has no coherent definition. A spreadsheet macro eliminates more clerical work than most industrial arms. The more defensible version is simply to stop preferentially subsidizing capital equipment relative to payroll.

Education and Retraining

The default political answer, and the one with the weakest evidence behind it. Rigorous evaluations of displaced-worker retraining programs consistently show modest effects. Retraining a 55-year-old logistics coordinator into a data analyst is a much harder problem than the policy language implies, and the honest record is that most such programs do not do it.

This does not mean skills investment is worthless. It means it is not a substitute for income support, and treating it as one has been the default failure of transition policy for forty years.

Second-Order Impacts

Health insurance decouples or collapses. In the United States, employment carries health coverage. Widespread displacement therefore produces a health crisis as directly as an income crisis, and the pressure to sever that link may end up being automation's most consequential political effect.

Geography concentrates the pain. Automation lands unevenly on the map. Areas dependent on a single automatable industry lose employment, tax base, and services at the same time, and the resulting decline is self-reinforcing. The political consequences of the last such round are still unfolding.

The career ladder loses its rungs. Junior roles are the most automatable, and junior roles are how professions produce senior practitioners. Firms optimizing headcount today are consuming a training pipeline whose replacement cost will not appear for a decade.

Bargaining power shifts before employment does. Workers do not need to be replaced to be weakened; the credible possibility of replacement is sufficient to suppress wages. This effect arrives earlier than displacement and is much harder to measure.

Care work becomes the largest remaining sector. The work least susceptible to automation is care, and it is currently among the worst-paid and least-respected. If it becomes the economy's largest employment category, its compensation and status become a central political question rather than a marginal one.

Beyond Economics: Work and Meaning

Work supplies things wages do not. It structures time. It supplies involuntary social contact with people outside one's chosen circle—historically one of the few forces mixing a society across class and politics. It supplies a legible answer to what a person is.

The evidence on this is unusually clear. Unemployment reduces life satisfaction substantially more than the associated income loss explains, and the effect persists rather than adapting away.² Whatever employment provides, money is not all of it.

Two cautions complicate the pessimistic reading.

First, retirement exists and is mostly fine. Hundreds of millions of people have left work without collapsing, largely because retirement is socially legitimate, financially provided for, and shared with a peer cohort. Voluntary, funded, expected non-work looks very different from involuntary, unfunded, stigmatized non-work. The difference is not leisure. It is legitimacy.

Second, the meaning defense assumes current work supplies meaning, which for many people it does not. David Graeber's argument about jobs whose holders privately believe them to be pointless suggests the meaning crisis may already be here, concealed by the fact that these jobs pay.¹⁰ If a substantial fraction of employment is already experienced as purposeless, then automating it removes income without removing meaning, and the policy problem is narrower than it appears.

The Path Forward

Near-Term Likely (2026–2032)

Augmentation dominates. Most workers use AI to produce more rather than being displaced by it.

Displacement concentrates in specific roles rather than spreading evenly: tier-one support, data entry, routine content production, first-pass document review, junior analysis. Entry-level hiring in affected professions contracts before headcount does.

Wage pressure builds at the median. Productivity gains flow disproportionately to capital and to the workers who direct AI systems, while employers discover that one person with good tools does what three did. The historical pattern in which productivity eventually lifted all wages depended on labor scarcity that may not hold when the substitute is general-purpose. Workers whose value rests on physical presence, interpersonal judgment, or creative direction retain pricing power longest.

Retraining programs are funded and produce mixed results, consistent with their forty-year record. UBI and job guarantees are discussed extensively and implemented nowhere at scale.

Plausible (2032–2040)

Knowledge work is meaningfully affected in aggregate rather than at the margins, and white-collar displacement becomes politically salient in a way blue-collar displacement was allowed not to be.

New occupations emerge and are, as always, unpredictable in advance. Whether they emerge fast enough is the open question.

Some form of income support beyond traditional welfare is enacted somewhere in the developed world—most likely a negative-income-tax variant rather than true UBI, and most likely in a small, high-trust country first.

Health insurance is severed from employment in at least one major economy, driven by displacement rather than by health policy.

Wild Trajectory (2040+)

Abundance makes basic provision cheap enough that income support ceases to be fiscally contentious, and employment becomes genuinely optional for a large fraction of the population.

Or: displacement outruns policy, gains concentrate, and the result is durable mass economic irrelevance for a substantial minority—the outcome that produces political instability rather than post-scarcity.

Or, most likely: neither. A long, uneven muddle in which some sectors hollow out, some regions never recover, patchwork policy responses partially work, and the aggregate statistics look tolerable while specific populations are quietly abandoned. This has been the historical pattern, and there is no strong reason to expect a cleaner one.

Risks and Guardrails

Displacement outpacing adjustment. The guardrails are countercyclical: portable benefits decoupled from employer, wage insurance covering the gap when displaced workers take lower-paying work, and extended transition support measured in years rather than weeks.

Gains concentrating in capital. Guardrails: taxing capital income comparably to labor income; broadening ownership through sovereign wealth funds or employee ownership so that returns to automation reach households that do not currently hold equity; antitrust enforcement preventing the productivity gains from being captured as monopoly rents.

Erosion of bargaining power. Guardrails: sectoral bargaining that covers workers regardless of employer; enforcement against misclassification; the recognition that labor market power, not just labor market participation, determines wages.

Regional collapse. Guardrails: place-based investment rather than relocation assistance alone; maintaining public services through the tax-base decline rather than after it; treating regional economic failure as a predictable consequence rather than a surprise.

Loss of legitimacy for non-work. Guardrails: this is cultural rather than fiscal, and it is the one most likely to be neglected. If a large population is going to spend less time in paid employment, the difference between that being retirement and that being unemployment is whether society treats it as legitimate. That is a matter of narrative, status, and institutional design—and it is cheaper to build than any of the transfer programs, while probably mattering more.

The Deeper Questions

Is work good, or did necessity teach people to say so? For most of human history work was toil and the aspiration was to escape it; the elevation of work into a source of virtue and identity is historically recent and culturally specific. It is genuinely unclear whether that is a discovery about human nature or a rationalization of an unavoidable condition. AI may be the experiment that finds out.

What is owed, and by whom? If AI-driven abundance is built on accumulated human knowledge—every text, image, and piece of code that trained these systems—then the claim that its returns belong solely to the firms that assembled the training runs is at least contestable. How that claim gets resolved determines who the gains reach.

What does a good life look like without a job at its center? The oldest question in philosophy, made suddenly operational. Societies that have faced a version of it—through aristocracy, monasticism, or mass retirement—produced answers, and those answers involved structure, obligation, and community rather than unlimited free time.

Conclusion

The question this chapter opened with has no technical answer.

Whether AI-driven productivity produces broad prosperity or concentrated wealth and stranded populations is not determined by how capable the models get. It is determined by tax policy, labor law, benefit design, antitrust enforcement, and whether health insurance stays attached to employment. Those are choices, and unlike the technology, they are choices that can be made deliberately and reversed if wrong.

The historical record on making them well is mixed at best. The prosperity of the postwar decades was not an automatic consequence of postwar productivity; it required union density, progressive taxation, public investment, and a political settlement that has since been substantially dismantled. Where those arrangements were weaker, the same productivity produced considerably less broadly shared benefit.

What the record does establish is that the adjustment mechanisms are slow. Institutional responses to previous automation waves arrived a generation after the displacement did, which was tolerable when the displacement took a generation too.

That is the specific reason to start now on a transition that has not yet arrived. Not because mass unemployment is imminent—it is not—but because the policy apparatus that would cushion it takes about a decade to build, and it has to exist before it is needed rather than after.

The technology is arriving on its own schedule. The response has to be built on ours.

Endnotes — Chapter 57

1. McKinsey Global Institute (2023) estimates that roughly 30 percent of hours worked in the US could be automated by 2030 using currently available technology. This is an estimate of automatable hours, not of job losses—a distinction routinely lost in reporting.
2. Research on unemployment and wellbeing by Andrew Clark, Andrew Oswald, and others shows that unemployment reduces life satisfaction well beyond what the income loss explains, and that people do not fully adapt to it over time.
3. Finland basic income pilot (2017–2018): €560 per month to 2,000 unemployed people. Results showed reduced stress and improved wellbeing, no reduction in job-seeking, and a small increase in employment.
4. GiveDirectly Kenya: the largest and longest-running cash transfer trial, providing roughly $22 per month for twelve years to more than 20,000 people. Results show sustained welfare improvements without evidence of reduced work effort.
5. US labor force participation for prime-age men declined from roughly 97 percent (1954) to about 89 percent (2024)—a structural withdrawal that long predates modern AI and does not register in headline unemployment figures.
6. The Earned Income Tax Credit is the largest US anti-poverty program for working households, providing up to approximately $7,000 per year to qualifying families, with well-documented effects on employment and child poverty.
7. Negative income tax: proposed by Milton Friedman; provides a guaranteed floor that phases out as earnings rise. Structurally similar to the EITC and substantially cheaper than a universal payment.
8. India's MGNREGA guarantees 100 days of wage employment per year to rural households and employs more than 50 million people annually—the largest job guarantee program ever implemented.
9. Historical automation transitions: US agricultural employment fell from roughly 40 percent of the workforce to about 2 percent over the twentieth century, and manufacturing from roughly 25 percent to under 10 percent. New sectors absorbed the displaced, over generations rather than years.
10. David Graeber, Bullshit Jobs (2018), argues that a substantial share of modern employment is experienced as pointless by the people performing it—suggesting that a crisis of meaning in work may precede rather than follow automation.