Every automation panic since the Luddites proved wrong about jobs — is AI different?

You watch the tool do a piece of your job. You describe what you want in a sentence or two and a paragraph comes back that is roughly what you would have written, or a block of code that runs, or a summary of a document you had not finished reading. It is not perfect, and you can see where it is thin, but it did in seconds a thing that would have taken you the better part of an hour, and the thought arrives on its own: if it can do this part, how far behind is the rest? This chapter is about the first half of that worry, the one about your own job and jobs like it. The larger half, what becomes of an economy if machines can eventually do most of what people are paid for, how a society would share out income once work is no longer the main way it is earned, belongs to the Frameworks volume and its chapter on which economic futures are actually plausible, and is not retold here. The question here is narrower and older than it looks: when a machine takes over part of the work, what happens to the people who were doing it?

The alarm, and its record

The fear is not new, and neither is the shape of it. In 1811 and 1812 the textile workers of the English Midlands, the original Luddites, broke the frames and looms that were doing in a fraction of the time what a skilled weaver did in a day. They were not confused about machinery; they were watching a specific and hard-won skill lose its price, and they were right about that part. A little over a century later, in the Depression’s first year, John Maynard Keynes gave the fear a name, warning of a “technological unemployment” that arrives when we discover ways of economizing on labor faster than we find new uses for it. The alarm returned with the automation scares of the 1950s and 1960s, when a presidential commission was empaneled to study what the assembly-line robot and the mainframe computer would do to employment. The economist David Autor, reviewing this long record, notes how reliably the commentary of each era overstated how much of human work the new machines would swallow and understated the ways the machines would also make human labor more valuable.

Set against the two centuries that followed the first looms, the strong version of the fear, that machines would leave most people with no work to do, has been wrong every time. Mechanization did not hollow out employment; across a run of transformations that each destroyed whole occupations, the feared permanent mass of workers with nothing left to do never formed: across the twentieth century, unemployment shows no long-term rise, and the share of the population in work rose rather than fell. The weavers were displaced and their particular craft did not come back, but weaving was not the last job, and the economy that put looms out of human hands went on to employ far more people at tasks the loom-breakers could not have named. That much of the reassuring slogan the record supports: on the question of whether automation ends work, the record of two hundred years runs one way.

What the slogan leaves out is the middle. The end of the story was kind to employment; the beginning of it was not kind to the first people who lived through it, and the gap between those two facts is what this chapter is about.

Wrong in the end, right for a generation

A closely studied case of that gap is the one the economic historian Robert Allen examined in the British industrial revolution, and named, after the young Friedrich Engels who described it, the Engels’ pause.

Read the two lines together. After 1800 they part: output per worker pulled steadily ahead, the real wage stood nearly still for thirty years and then lagged for thirty more, and the gap did not begin to close until after 1860. On Allen’s figures, output per worker rose by close to a half between the 1780s and 1840, while the real wage rose by little more than a tenth: a generation of workers produced far more and were paid barely more, and the difference went to the profits that financed the mills and railways. Then the pattern turned. Between 1840 and 1900 wages grew faster than output and made up the lost ground, arriving at the modern arrangement in which pay and productivity advance roughly together. The gains were real and, in the end, widely shared. They were also, for the people who lived the first half of that century, a long time coming.

How deep the pause ran is a genuine dispute among historians, and the figure shows it as one. Allen builds his wage line on Charles Feinstein’s estimates; Gregory Clark has argued that real wages grew faster, and output more slowly, than those series say, which would make the pause shallower and the early gains less lopsided, and in Clark’s reading it is workers rather than owners who come out ahead in the period. Allen’s answer, set out in the paper’s appendix, is that Clark’s revisions do not hold up against the tax and income records, and that even substituting Clark’s own wage series leaves a pause in place, shallower but present. The two accounts disagree about the depth of the wage lag and agree that there was one. For the argument of this chapter, the existence is what matters: the machines that eventually enriched Britain first delivered a long stretch in which output soared and the typical wage did not.

Why the alarms were wrong in the end

If displacement is real and sometimes lasts a generation, why did employment always recover rather than collapse? The answer has two parts, and both come from thinking about jobs as bundles of tasks rather than indivisible things. When a machine takes over a task, it does two things at once. It substitutes for the worker who did that task, which is the part everyone sees. It also, less visibly, raises the value of the tasks the machine cannot do, because those tasks now stand between the cheap, automated step and a finished product someone wants. Autor calls this the complementarity that the commentary of each era tends to ignore, and in a 2015 essay, drawing on work by James Bessen, he offers a modern example: the automated teller machine might have been expected to abolish the bank teller, yet teller employment rose a little over the three decades to 2010, even as tellers shrank as a share of all work, because the machines made branches cheaper to run, banks opened more of them, and the tellers who remained spent less of their day counting cash and more of it selling services and advising customers. The tasks a machine leaves alone become the more valuable half of the job.

The second part is what the cheaper output does downstream. When automation makes goods and services cost less to produce, the money not spent on them is spent on something else, and the something else is made by people. Rising productivity has, over two centuries, shown up as rising real incomes, and rising incomes buy goods and experiences that did not exist or were luxuries before, from air travel to physical therapy to restaurant meals, each of them a source of work. The displaced weaver’s grandchildren were not re-employed weaving; they were employed in trades the weaver had no word for. Reabsorption of this kind is not a law of nature, and it is not instantaneous, as the pause shows. But it is the mechanism by which two centuries of labor-saving invention failed to produce the permanent joblessness each era feared, and the question cuts both ways: a claim that this time is different has to explain why the mechanism would now fail, and a claim that history guarantees a repeat has to explain why the mechanism would survive a change in which tasks are exposed.

What made the mechanism work was that the tasks machines could not touch were broad and human. Autor draws on an observation of the philosopher Michael Polanyi, who noted that we can know more than we can tell: a great deal of human skill, from recognizing a face to calming a frightened patient to knowing which way a negotiation is turning, rests on tacit knowledge that the person applying it could not write down as a rule. Machines that had to be programmed with explicit rules could not do such tasks at all, so those tasks became the cushion, the part of human work that stayed valuable and absorbed the people whose codifiable tasks had been automated away.

Is this time different?

That is the question the newest wave puts on the table, and it is a real question rather than a settled one, because the thing that made the old reassurance work is exactly what the new machines are built to change. The cushion held because tacit, non-codifiable tasks were beyond the reach of machines that needed explicit rules. Contemporary computer science, in Autor’s words, “seeks to overcome Polanyi’s paradox by building machines that learn from human examples”: rather than being handed rules, such systems infer the patterns behind those examples, which lets them attempt precisely the tasks that could never be reduced to rules. A system that writes a passable first draft, or fixes a bug, or reads a contract for the risky clause, is reaching into the cognitive, non-routine tasks that were the cushion in every previous wave.

What differs from earlier automation, then, is which tasks fall on each side of the line, not whether tasks are exposed at all; some exposure was always there, and what has moved is where the line runs.

How far that reach extends is the subject of a young and fast-moving literature, and its most-cited early estimate is explicit about what it is measuring and what it is not. Studying the tasks that make up American jobs, a team led by Tyna Eloundou estimated that around 80 percent of workers have at least a tenth of their tasks exposed to large language models, and roughly 19 percent have at least half their tasks exposed, with higher-paid jobs tending to be more exposed rather than less, a reversal of the pattern of earlier automation. The authors are explicit that exposure is not displacement: their measure asks whether a task could be done faster with the technology, not whether a job will be lost, and they state that they make no prediction about how fast the tools will be built out or adopted. Taken at its own word, the study reports that the reach is broad and lands differently than before, and it declines to say what the reach will do. That is the state of the exposure evidence, and it is incomplete rather than only contested.

The first studies of outcomes, rather than exposure, began arriving in 2025 and 2026, and they read as early rather than decisive. In Denmark, where by their survey most employers in exposed lines of work had taken the tools up, Anders Humlum and Emilie Vestergaard matched adoption surveys to administrative records and found the measured effect on earnings and hours, two years in, close to zero, with anything larger than a couple of percent ruled out, even as employers reorganized tasks around the tools. A team led by Menaka Hampole, tracking American firms, found demand falling for the tasks most exposed while overall employment at the adopting firms held up, the loss in exposed work offset by the extra demand that higher productivity created in the same firms. And in the Census Bureau’s large national survey of business AI use in early 2026, two thirds of adopting firms reported using the tools to help workers with tasks rather than to replace them, and outright employment reductions were rare. Each of these covers a short window on an early stage of adoption, and each says so; none measures what a decade of buildout would do. They are partial answers, and what they answer so far is the short run.

So the reassurance and the alarm both overreach the same evidence from opposite sides. Whether the reabsorption that worked for two centuries works again turns on questions that are open: whether the tasks language models cannot do are broad enough to be the next cushion, whether the complementarity that made automated tasks raise the value of human ones holds when the automated tasks are cognitive, and how fast the change arrives relative to how fast workers and firms can adjust. None of these is settled yet by the evidence. Where the displaced people land, and whether the places they live recover, is a question this book takes up in the Countries volume, in the chapter on why some regions never recover after their industry dies; what displacement costs a person beyond the wage is the subject of the next chapter, on work and purpose.

Neither reassurance nor alarm

The two-century record and the newest evidence point in directions that a careful reader has to hold at once. Automation has not ended work, and the mechanism by which it kept failing to end work, cheaper output feeding new demand and untouched tasks rising in value, is real and carried two centuries of change. That is a reason not to reach for the alarm that this time the machines take everything. At the same time, the cushion that made the mechanism work was a particular set of tasks that the newest systems are aimed squarely at, and the studies measuring their reach state that they are measuring exposure and not outcomes, while the earliest studies of outcomes cover only the opening years of adoption. That is a reason not to reach for the reassurance that history guarantees a soft landing. The pattern was wrong in the end and right for a generation, and whether it holds a third time is not something the present evidence can settle. What can be said is that if the pattern does repeat, the cost will again fall in the middle, on identifiable people and places, and that the size and shape of that cost is the part worth watching, whichever way the larger question turns out.

The next chapter stays with those people rather than the machines. If a job can be lost, and the loss cuts deeper than the paycheck, then work is doing something for a person beyond paying them, and it is worth asking what that something is, whether the loss of it is what makes unemployment hurt, and whether it can be had another way. The chapter turns to work and purpose, and to whether people fall apart without a job.