The Machine

Between 1811 and 1816, in the textile districts of England, bands of workers broke into mills by night and smashed the power looms and stocking frames that were taking their livelihoods. They called themselves Luddites, and they have lent their name ever since to a fear as old as machinery itself: that the machine will do the work, and the worker will be left with nothing. A century later Keynes gave the fear a tidier name — “technological unemployment” — though he judged it a passing ailment, the growing pain of an economy learning to need fewer hands. The name and the fear have returned with every great wave of mechanisation since: the assembly line, the computer, the robot, and now the thinking machine.¹

The fear has a simple and forceful logic. A wage is paid for work; if a machine can do the work more cheaply than a person, the person is redundant, and the wage falls to nothing. Multiply that across enough tasks and the conclusion is mass unemployment, a world in which the machines do everything and the workers are simply surplus. Every previous wave was answered by the comforting observation that displaced workers found new work — but that comfort rested on the existence of higher ground, tasks the machines could not yet reach. What gives the present moment its edge is that the machine has, for the first time, climbed onto the cognitive ground itself: it writes, it codes, it reasons, it draws. The places to which displaced hands always retreated — the work of the mind — are now part of what is exposed.

And yet the Luddite has been wrong for two hundred years, and spectacularly so. Through two centuries of relentless automation, employment did not collapse and wages did not fall to nothing; real wages rose many times over, and most people work less and have far more than their great-great-grandparents could have dreamed. The pessimist's error has a name: the lump-of-labour fallacy, the assumption that there is a fixed quantity of work to be done, so that every task handed to a machine is a task stolen from a human. The quantity of work is not fixed, and three forces see to it.²

The first is income. A machine that does a task more cheaply lowers the price of what it makes, which leaves people with more to spend on everything else — and that spending becomes demand for other labour, often in industries that did not exist before. The second is complementarity. A machine that takes over part of a job can make the workers who do the rest of it more valuable, not less. The classic case is the automated teller: as cash machines spread through the 1970s and after, the number of bank tellers did not fall but rose, because the machines made each branch cheaper to run, so banks opened more branches, and the tellers turned to the work the machines could not do — selling, advising, solving. The third is invention. Technology conjures entirely new kinds of work; a great share of what people are paid to do today are jobs that did not exist a hundred years ago. Between them, these forces have always more than replaced what the machines destroyed.³

But the optimist's victory is an average, and an average can hide a multitude. That real wages rose across two centuries does not mean every worker prospered; it is consistent with the loom-breaker ending his days in poverty while the factory that replaced him enriched a nation. To see who wins and who loses, and when, the binary of pessimist and optimist must give way to something more exact — the task-based view of technology, the framework that now organises the field.⁴

Think of a job not as a single thing but as a bundle of tasks. Technology acts on the wage by acting on tasks, and it pushes in two directions at once. On one side is displacement: a machine takes over a task a worker used to do, and the demand for that worker — and the wage it commanded — falls. On the other side are two countervailing forces, which we may gather under the name reinstatement: the productivity the machine adds, which lowers prices and swells demand, and the new tasks it calls into being, tasks where human labour holds the advantage. Displacement pushes the wage down; reinstatement and productivity push it up. What happens to the wage is the outcome of the race between them.

This is why technology has no fixed sign for the wage. When reinstatement and productivity outrun displacement — when each machine that takes a task also helps create new and better-paid work — wages rise, and the optimist is vindicated. When automation is what Acemoglu and Restrepo call “so-so” — just good enough to replace a worker, but not good enough to generate much new productivity or much new work — displacement runs ahead, and wages stall or fall. The self-checkout kiosk that barely outperforms the cashier it replaces is the type case: it displaces without reinstating. The question is never simply whether there is automation, but whether the renewal keeps pace with the loss.⁵

The race explains the history. For the few decades after the Second World War, automation was rapid and wages rose broadly across the whole distribution; the machines of that era displaced some tasks but reinstated handsomely, and the gains were widely shared. Then, from around 1980, the pattern changed. The new wave of automation fell with particular force on routine tasks — the predictable, codifiable work of the clerk, the bookkeeper, the assembly operative — because routine is precisely what a machine does best. And the reinstatement that followed did not lift the same workers.

The result is the phenomenon labour economists call polarisation, or more vividly the hollowing-out of the middle. Employment and wages grew at the two ends of the skill ladder and shrank in the centre, and the routine jobs that had been the heart of a broad working and middle class were handed to software and machines.⁶ There is a warning in the recent record. The reinstatement that rescued earlier generations has, of late, been running behind: the new work is being created more slowly than the old work is being automated away, and the wage of the ordinary worker has paid the price. The optimist's two-century winning streak is not a law of nature. It held because renewal kept pace — and renewal can fall behind.⁷

Which brings us to the machine now reshaping the argument. Artificial intelligence — and in particular the generative models that have spread since the early 2020s — differs from its predecessors in a way that matters. Earlier automation took over manual and routine tasks, leaving the cognitive ground to humans. Generative AI is cognitive: it writes and edits, codes and debugs, translates, summarises, analyses, draws. The exposed tasks are now, for the first time, the very ones to which displaced workers have always climbed.⁸ The hard question is whether this makes AI a displacing technology, hollowing the cognitive middle as software hollowed the clerical one — or a reinstating one, complementing human workers and making them more productive.

The early evidence, which must be read as a first sketch and not a verdict, is genuinely mixed, and it cuts both ways. On the hopeful side, the productivity studies are striking: given an AI assistant, customer-service agents, writers, and programmers have all grown measurably more productive — and, tellingly, the largest gains accrue to the least experienced workers, the tool lifting the bottom of the skill distribution toward the top. In those settings AI looks strongly complementary, a machine that augments rather than replaces, and that narrows rather than widens the gap between workers.⁹ On the worrying side, in the online markets where freelancers sell exactly the tasks AI does well — writing, coding, translation, image-making — demand has visibly fallen, and the work that vanishes first is the entry-level work by which novices once climbed. Across the broad economy, through 2025, careful studies find no aggregate employment loss from AI — adoption has reached roughly a third of workers, and wages in exposed jobs have if anything edged up — but they also find the disruption concentrated on the young and the just-starting, those with the least power to absorb it.¹⁰

The honest position is humility. The optimist's record is long and should not be lightly dismissed; so far, the aggregate numbers are calm. But the cognitive reach of the new machine is real, the pace is fast, and the warning from the recent decades — reinstatement falling behind — is exactly the condition under which the optimist's streak would end. We do not yet know whether AI will reinstate as generously as it displaces. What we can say is that the answer will not be dictated by the technology alone — and that is the subject of the rest of this plate.

Return now to the frame this volume has carried throughout: the wage as a bargain settled in a range, between a ceiling set by what the work is worth and a floor set by the worker's alternatives, with power deciding where in that range the wage falls. Technology, it turns out, reaches into this picture twice — once on the side of worth, and once, less obviously, on the side of power.

Its action on worth is the one the whole discussion so far has described. By displacing labour from some tasks and complementing it in others, technology changes what work is worth — lowering the ceiling where machines substitute, raising it where they assist — and, on the whole and over time, lifting the total value produced. If that were the only channel, the dispute would reduce to the race already described, and a society that grew more productive would, sooner or later, grow richer in wages too.

But there is a second channel, and it is the one the volume's argument insists upon. The machine is the employer's ultimate outside option. Recall that the worker's position in the bargaining range depends on the employer's alternatives as much as the worker's own: a worker who is easy to replace has little power, whatever their worth. A machine that could do the job is the most complete replacement of all — and here is the crucial point, the machine need not actually be installed to do its work in the bargain. Its mere availability, the credible prospect of it, is enough to cap the wage. An employer will not pay a worker much more than it would cost to automate the task; the cost of the machine becomes a ceiling on the wage, pressing down from above.

And that ceiling descends. As machines grow cheaper and more capable, the cost of replacement falls, and the cap it places on the wage falls with it — even for the worker who is never replaced, who keeps the job and watches the wage it pays slip downward toward the price of the machine that shadows it. This is the deep reason productivity and wages can part ways. Technology can raise what the work is worth — can lift the ceiling of true value — while at the same time arming the employer with a cheaper and cheaper threat, which lowers the worker's power to claim that value. The pie grows, and the worker's slice shrinks, and there is no contradiction: the first is the work of worth, the second the work of power. The widening gap between what workers produce and what the typical worker is paid, which opened in the rich economies around 1980, is this divorce made visible — and the next plate takes it up as the volume's final question.¹¹¹²

So the future of the wage is not, at bottom, a question about machines. It is a question about who keeps what the machines make. Technology sets the size of the prize — and, the pessimist's recurring error aside, it has made the prize enormous and will likely make it larger. But the division of that prize between those who work and those who own is not settled by the technology. It is settled, as everything in this volume has been settled, by power: by whether workers have the alternatives, the organisation, the legal floors, and the ownership stakes that let them claim a share of the value, or whether that value flows by default to whoever holds the machines.

This reframes the whole anxious debate about automation and artificial intelligence. The fear that the machines will take the jobs and the reassurance that new jobs will appear are arguing about the size of the prize — and on that question the optimist usually wins. But both are missing the question that actually decides the wage, which is the distribution of the prize, and that question the technology does not answer. A society in which workers are strong can turn a wave of automation into shorter hours, higher pay, and better work, the productivity shared out. A society in which workers are weak can turn the same wave into displacement, falling wages, and fortunes concentrated among owners, the productivity captured. Same machines; opposite outcomes; and what divides them is not engineering but the balance of power. The instruments the earlier plates described — the union, the minimum wage, the safety net, the broad ownership of skill and capital — are precisely the means by which a society decides that the gains of the machine will be shared. They are not obstacles to progress. They are the difference between progress that lifts the worker and progress that leaves the worker behind.¹³

The Luddites, then, were wrong in the way everyone remembers and right in a way usually forgotten. They were wrong that the loom would impoverish the nation; it helped enrich it beyond their imagining. But they were not wrong that it could impoverish them — the particular weavers whose skill it made worthless, who were not among those who kept the gains. Whether the thinking machines of this century enrich the workers whose tasks they take, or only the owners who deploy them, is the same open question, and it has the same kind of answer. It will be decided not in the laboratory but in the same long contest over power that this whole volume has traced.