The Machine

The machine can enrich a country while impoverishing the worker whose task it takes. The two events can belong to the same history, and the interval between them can last the rest of that worker's life. This is the part of the old fear that an account of rising national prosperity cannot answer on its own. A person is paid to perform particular work, in a particular place, with capacities acquired over years. A cheaper way of doing that work changes the terms on which the person can sell it. Somewhere else a new occupation may be forming, a cheaper product may be drawing new customers, or another worker may be becoming more valuable beside the equipment. None of those developments gives the displaced worker an immediate claim on the new income. The question for this plate is accordingly the machine's passage through the wage: how a saving of labor in production changes the demand for a person's work, the alternatives within reach, and the share of a growing product that the person can keep.

Think of a job as a bundle of tasks. A machine can take over one part of the bundle while increasing the value of the person who performs the rest. Calculation can become cheaper while judgment about what to calculate becomes more useful; routine processing can accelerate while an unusual case still needs someone who understands the customer. Complementarity has a demand side as well. If the saving lowers the price and expands the market enough, a firm may need more of the remaining human work even while using less of it for each unit sold. But this is a mechanism whose strength must be established, not an assurance attached to the word productivity. A market that barely expands can release workers; a machine that takes over most of the valuable tasks can leave little for its former operator to supply. David Autor's account of automation places substitution and complementarity in the same job, making the composition of work central to what happens to its price. The surviving task matters because the wage belongs to the person who can perform it, rather than to the job title left on the payroll.¹

The task framework separates three movements that an argument about the machine can too easily run together. Displacement occurs when capital takes over tasks previously assigned to labor. Productivity gains make production cheaper and can expand the demand for the tasks labor still performs. Reinstatement occurs when new tasks are created in which labor has a comparative advantage, adding work to labor's side of the boundary. The last two can counter the first, but they are different sources of a recovery in labor demand. A machine can save enough on a narrow operation to be adopted while delivering only a modest gain in overall productivity; the displaced work can then outweigh the additional demand generated by the saving. Acemoglu and Restrepo give this possibility its useful name, so-so automation. A transformation that looks impressive inside the plant need not raise the demand for workers outside it. Nor does the net movement in labor demand by itself decide how much appears as wages and how much as employment. That division also depends on labor supply and on the way the labor market sets its terms.²

Even a strong recovery in the demand for labor leaves the identity of the worker unsettled. The person released from one task may lack the skill, location, time, or money needed for the task that replaces it. Experience valuable in the old occupation can lose its market before it has finished earning its return. A young entrant can train for the new work; a worker with dependents and a mortgage must finance the passage while continuing to live. These are costs of adjustment borne by people, and an aggregate count of jobs can conceal them. The preceding plates on skill and insurance therefore return inside the machinery. Training affects who can supply the new complement; support during unemployment affects whether someone can search and retrain rather than accept the first available wage. The same technological change can also divide a workforce, raising the demand for one set of capacities while reducing the demand for another. Autor's account of the changing task content of work gives the distributional question its proper scale: which activities are displaced, which become more valuable, and which workers can make the passage between them.³

A machine that assists with language makes this distinction newly visible without abolishing it. In a study published in 2025, Erik Brynjolfsson, Danielle Li, and Lindsey Raymond examined a generative-AI assistant used by 5,172 customer-support agents. Access increased issues resolved per hour by about 15 percent on average, with larger gains among less experienced and lower-skilled workers. The tool could help a novice use patterns of response that had previously depended more heavily on experience. The study followed a staggered rollout within one firm, using comparisons across agents and time to estimate the effect. It did not observe the resulting changes in wages or aggregate labor demand. A gain in an agent's measured output therefore supplies one part of the wage inquiry: the tool altered what the agent could do in an hour. Whether the firm expanded service, reduced staffing elsewhere, raised pay, lowered prices, or retained a larger margin belongs to the further inquiry into adjustment and distribution. The experiment in work is concrete; a prediction about the income of every worker would require evidence beyond it.⁴

The machine also enters a bargain before it enters the factory. Where the employer and worker divide a surplus, a credible alternative to continuing the existing arrangement changes what either side is willing to accept. Automation can improve the employer's alternative, just as a competing offer can improve the worker's. But the price on a machine is not a universal ceiling on the wage. Installation takes capital and time; maintenance, reliability, supervision, and the remaining human tasks enter the comparison. A machine that performs one operation does not necessarily replace the person who performs a dozen. A threat that would cost more to carry out than it saves has a different force from an available replacement the firm would actually choose. These details matter because the wage responds to the value of the whole alternative, not to a demonstration of one technical capability. The earlier plates on bargaining and monopsony supply the setting for this inference. They also show why an increase in the worker's productive capacity need not pass into pay in the same proportion: the capacity and the terms on which its use is purchased can change together, in different directions.⁵

Whose machine, then? The question includes who owns the equipment, who can use it, who acquires the complementary skill, and who bears the cost of becoming unnecessary to its operation. Technology changes both what an economy can produce and the relative demand for the contributions by which it produces it. Institutions govern the claims made within that changing economy: the wage floor, collective bargaining, insurance, training, and the distribution of ownership can all affect who receives the gains and who finances the passage. Shorter hours, better work, higher compensation, lower prices, and larger profits are distinct possible uses of a productive saving; their division requires an account of the market and the bargain through which the saving travels. The original fear returns here in its strongest form. A worker need not deny the value of the machine to ask for a place in the prosperity it creates. The last plate gathers those claims across the economy, adding the wages together and asking what part of the whole has reached the people whose work remains inside it.