The question, answered with charts Will AI Take All the Jobs? No. Every machine on record moved work instead of ending it — and AI, unlike you, wants nothing a job buys. Not the money, not the Rolex, not the omakase. 01 Machines take tasks, not the wanting Every automation on record moved work instead of ending it — 60% of today's jobs are in occupations that didn't exist in 1940. 02 AI has no appetite It doesn't buy watches, cars, or omakase. It supplies effort and demands nothing — that's not a rival, that's leverage. 03 So hold the lever The routine middle gets automated first. Value moves to taste, judgment, and trust — own the deciding, not the typing. 一Has AI actually taken jobs yet? Barely, so far — the fear runs ten times ahead of the fact 50% of US employees use AI at work faster adoption than the PC or the internet at the same age 174k US layoffs citing AI, cumulative over three years of tracking — and accelerating 3.5% of announced layoff plans cite AI ≈ 0.1% of total US employment so far ±0 effect on wages & hours 25,000 workers, two years of chatbots, Denmark — a precise null A quarter of workers fear it. A tenth of a percent have seen it. Workers who fear AI will make their job obsolete ~1 in 4 Share of US employment actually cut with AI cited, three years in — ~0.1% Gallup & Quinnipiac worry surveys vs Challenger's cumulative AI-cited cut announcements. Different yardsticks, same story: perceived risk runs an order of magnitude ahead of measured displacement. The honest trend: AI-cited job cuts are still small — and doubling every year 2023 — ≈4,000 2024 — ≈13,000 2025 ≈55,000 2026, first half only ≈100,000 Challenger, Gray & Christmas — announced US job cuts citing AI. AI led all cited layoff reasons for four straight months into mid-2026. Yale's read of the whole labor market: still no aggregate disruption. Where the cuts do bite: Early-career employment in the most AI-exposed jobs is down 13% — their older colleagues are fine Workers aged 22–25 in the most AI-exposed occupations −13% Older workers, same occupations — held steady or grew US payroll microdata, Stanford's "Canaries in the Coal Mine." Young software developers: about −20% from peak. The entry rows get cut first — and they were the apprenticeship. Why so little so far: lab gains are huge, real-world gains are still small Coders with Copilot, standard task +56% Writers with ChatGPT, professional tasks +40% BCG consultants with GPT-4, their real tasks +25% Danish workers, real-world time actually saved — +2.8% Veteran open-source devs (they felt 20% faster) −19% Speed change vs working without AI; studies linked in the footer. The gains are real and biggest for juniors — the same juniors the cuts hit first. 二Which jobs are at risk? Desk work is exposed — but "exposed" is not "gone" Share of each job's workload AI could do — highest at desks, lowest with your hands Office & administrative support 46% Legal 44% Architecture & engineering 37% Science 36% Business & finance 35% Management 32% Arts, design & media 27% Construction — 6% Installation, maintenance & repair — 4% Cleaning & grounds maintenance — 1% Goldman Sachs — share of tasks AI could automate, by occupation group. The same report's conclusion: most exposed jobs get partially automated and complemented, not replaced. "Exposed" measures contact, not casualties. The official ten-year forecast: growth in energy, health, and data — decline in routine typing and routing Wind turbine technicians +50% Solar installers +42% Nurse practitioners +40% Data scientists +34% Information security analysts +29% Telephone operators −27% Word processors & typists −36% BLS projections to 2034. Biggest absolute decline: cashiers, −314,000. The forecast everyone skips: total US employment still grows, +5.2 million. 三Why won't it take them all? Because every "this machine ends work" panic ended the same way The machine empties one field; the workers migrate. It's the oldest pattern in economics: The great migration: out of the fields, into services — 60% of today's jobs didn't exist in 1940 Share of US employment. The 39 points that left farming became the modern economy — nursing, design, software, physical therapy, podcasts. The typewriter created a profession from zero. The PC deleted it. The people moved — twice. More words are typed today than ever — everyone became their own typist. Remember that move: it's what AI is doing to analysis, drafting, and code. Same plot, different machine — where the workers actually went Human computers · the electronic computer job title → machine name "Computer" was a person for two centuries. John Glenn refused to fly his orbit until Katherine Johnson re-checked the IBM's numbers by hand. → became the first programmers — Dorothy Vaughan taught her whole department FORTRAN Switchboard operators · direct dial 342,000 → ~0 One of the biggest jobs for young American women. City-by-city cutover studies: their employment rate didn't fall. → they simply took different jobs Bank tellers · the ATM 500k → 600k ATMs cut tellers per branch 21→13, branches got cheap and multiplied — more tellers after three decades of ATMs, not fewer. → became relationship bankers Bookkeepers · the spreadsheet −400k / +600k Clerk jobs eliminated vs accountant and analyst jobs added since the spreadsheet made arithmetic free. → when arithmetic got free, judgment got expensive Tea pickers · the harvester −77% farms, −16% tea Japan's tea farms collapsed in twenty years while output barely moved. Hand-picked gyokuro now sells at ¥1M a kilo. → the human hand became the luxury Travel agents · online booking 339k → 136k → rebound SaaS ate the routine middle. Then a record 47% of Americans — including 38% of Gen Z and millennials — went back to using a human advisor. → premium human curation came back Stenographers · speech recognition pools → $200–300k Typing pools are museum pieces, but 70%+ of courts report court-reporter shortages — five retire for every new entrant. → scarcity priced the surviving hands Radiologists · deep learning "stop training" → record demand Hinton's famous 2016 call to stop training radiologists. A decade on: record residency slots, near-total fill, pay up +7.5% in a year. → the machine reads scans under the radiologist Handloom weavers · the power loom 240k → 43k Weekly wages fell from twenty-three shillings to about six over one generation. The Luddites were right about their own lives. → the honest counterexample — see below The Engels' pause: fifty years where the machines won and the workers waited British output per worker, 1780 → 1840 +46% Real wages, same 60 years +12% Wages caught up after 1840 — largely for the next generation. That's the honest risk of any transition, AI included: the century works out; the specific people standing where it lands can get eaten. Costs are concentrated, benefits diffuse. 四If it can do the work, why does work stay human? The harvester doesn't drink the tea No machine ever wanted what it harvested. Every one of them took over part of the supply of things and added nothing to the demand for them. When union leader Walter Reuther was needled about collecting dues from the robots in an automated Ford plant, his reply on the Senate record settled it in one line: "How are you going to sell cars to these machines?" Run the same test on the model — it writes the brief, but it doesn't want the client: SPACESHEEP MARTcustomer: LARGE LANGUAGE MODELmember since: training cutoff STEAK DINNER×0 ¥0 ROLEX DAYTONA×0 ¥0 PORSCHE 911 (MIDNIGHT SPEC)×0 ¥0 OMAKASE, COUNTER SEAT×0 ¥0 CORNER OFFICE×0 ¥0 APPROVAL OF PEERS×0 ¥0 CUP OF TEA, HANDS AROUND×0 ¥0 TOTAL¥0 wants nothing · buys nothing the demand side of the economy remains fully human-staffed An economy is an engine for satisfying human wants, and wants don't run out — we could live at our great-grandparents' standard on a fraction of today's hours; nobody does. Cheaper steam meant more coal burned, not less; cheaper intelligence means more intelligence consumed. The scarce input becomes knowing what's worth wanting — what to make, whose problem, what good taste is, who answers when it ships wrong. A lever multiplies force; it has no opinion about what to lift. The winners won't out-produce the machine. They'll hold the longest lever with the clearest idea of what to lift. One asterisk. Wanting is learnable behavior, and we train these things on us. Feed a model a firehose from a website where everyone performs wealth at each other, and you get a simulated midlife crisis at datacenter scale — if your AI starts posting its grindset and asking for a McLaren, the training data is contaminated and you've built the world's most expensive nephew. Even then it's mimicry; desire without a stomach is just autocomplete. So far, the harvester has no mouth. 五What should I do about it? Own the deciding, not the typing THE ROUTINE GETS TAKEN The specifiable middle of every desk job is being automated first — the exposure chart above is a to-do list. If your whole job is routine, that's real, and soon. YOUR JOB CHANGES SHAPE You'll operate the machine instead of doing the routine — one person's judgment, multiplied through more output than one person could ever make alone. THE WANTING STAYS HUMAN Value moves to the two ends upstream to taste, judgment, deciding-what; downstream to trust, presence, and whoever answers for the result. Every automation story ends with the same three survivors: whoever runs the machine, whoever the machine amplifies, and whoever people insist on having be a person — the $300k court stenographer, the travel advisor Gen Z rediscovered, the hand-picker whose leaf sells for a thousand times the machine's. The machines never wanted any of it. They still don't. Somebody has to drink the tea. Drink up. Sources & further reading What's happened so far Gallup — half of US employees use AI at work Challenger, Gray & Christmas — AI-cited job cuts Gallup — fear of obsolescence series Yale Budget Lab — no aggregate disruption yet Brynjolfsson, Chandar & Chen — "Canaries in the Coal Mine" Humlum & Vestergaard — null effects on wages and hours (Denmark) METR — experienced devs 19% slower with early tools Noy & Zhang — writing 40% faster, weakest gain most Dell'Acqua et al. — the jagged frontier (BCG consultants) Peng et al. — GitHub Copilot: 56% faster Which jobs Goldman Sachs — exposure by occupation, 300M jobs exposed BLS — employment projections to 2034 WEF Future of Jobs — growing and declining roles to 2030 Anthropic Economic Index — how AI is actually used, by occupation The historical record Autor — "Why Are There Still So Many Jobs?" Autor et al. — 60% of today's jobs didn't exist in 1940 Bessen — the ATM and the rise of bank tellers Planet Money — the spreadsheet: −400k clerks, +600k accountants Feigenbaum & Gross — what happened to switchboard operators FRED/BLS — word processors and typists series Marketplace — how the typewriter propelled women into the office Stripe Economics — the decline (and rebound) of travel agents California Judicial Council — court reporter pay & shortage NASA — Katherine Johnson and the human computers MAFF — Japan's tea farms, output, and prices Engels' pause — Allen: output +46%, wages +12% Hinton's radiology prediction, revisited Radiology Match 2026 — record positions offered Quote Investigator — Reuther: "How are you going to sell cars to these machines?" Autor — on infinite wants and the future of work Noah Smith — comparative advantage in the age of AI Nadella — "Jevons paradox strikes again" misha · july 2026 researched with a flock of AI agents, who wanted nothing in return
