A machine can make a worker unnecessary without making work unnecessary. The distinction is easy to state from a distance and difficult to live through. A society may eventually produce more, employ people elsewhere and find uses for abilities it had not previously rewarded; the person whose task has disappeared still loses the particular wage on which his life was arranged. The question is therefore two questions. What happens to the demand for labor when a machine takes over a task? And what happens to the person who supplied it? An answer to the first does not settle the second. The gain can be real, the displacement can be real, and they need not arrive in the same place or at the same time.
The unit to follow is often the task rather than the occupation. A job joins activities that need not yield to the same technology: recording information, interpreting it, checking an exception, explaining a result, taking responsibility when the procedure fails. Automation can remove one activity and increase the usefulness of another. David Autor's account of this process emphasizes that machines substitute for some tasks while complementing others. A cheaper calculation can make judgment based on the calculation more valuable. But complementarity is not a promise made to everyone who used to calculate. The worker must be able to perform the remaining task, and employers must want enough of it. The composition of the job and the composition of the people seeking it can move at different speeds.
Consider a workshop in which a hundred finished units require a hundred hours of processing and a hundred hours of checking. A machine cuts processing labor in half while leaving checking unchanged. The same output now needs a hundred and fifty hours instead of two hundred. If lower costs expand sales to a hundred and twenty units, labor requirements become a hundred and eighty hours: output rises, productivity rises, and total hours still fall. If sales expand to a hundred and sixty units, hours become two hundred and forty. The machine has the same technical effect in both cases. What changes the employment result is how far demand expands. These are illustrative quantities, with constant task requirements and unchanged hours per worker; they contain no forecast about a particular industry.
Acemoglu and Restrepo distinguish this displacement from the effects that can counter it. Automation transfers an existing task from labor to capital. The resulting cost reduction can expand production and increase demand for labor in tasks that remain. New tasks in which people have an advantage can also appear, drawing labor into activities that did not previously form part of the production process. That last movement is different from doing more of the old work. It is reinstatement through a changed set of tasks. Whether these forces offset displacement is a question about their relative strength. A technology that removes substantial labor while delivering only a modest cost saving need not create enough additional demand to make the displaced hours reappear elsewhere.
Nor is the price on the machine's invoice the wage the worker must now accept. The relevant alternative is a usable production arrangement. Suppose a day's human production yields two hundred, requires forty in other costs, and pays the worker w. Its return to the employer is 160 − w. A machine delivering the same output costs a hundred and thirty a day once finance, installation, maintenance and the labor needed to operate it are counted. Its return is seventy. Under these assumptions, the employer prefers retaining the worker only at a wage no higher than ninety. A fifty-unit rental charge would not establish a wage ceiling of fifty if another eighty is needed to make the machine produce. The comparison concerns complete alternatives, not isolated prices.
Even the ninety in that example is a conditional ceiling for one relationship. If the worker can obtain eighty elsewhere, there is room for an agreement between eighty and ninety. If his alternative pays a hundred, that room disappears. He may leave for the other job rather than accept the ceiling. If the automated arrangement instead costs a hundred and ninety, the ceiling rises to a hundred and fifty, and a wage of a hundred is compatible with retaining him. If the human product is better, or the machine cannot handle the necessary exceptions, the supposedly equivalent alternative must be recalculated. A credible substitute can weaken bargaining power before it is installed. An unsuitable machine cannot acquire that power merely by being advertised.
This also resolves the apparent choice between the price of the machine and nothing at all. A worker who loses this job receives no wage from this employer. That is an employment outcome, not a statement that his labor has a zero price everywhere. Another employer may use the same skill; another occupation may require training; another region may be beyond a feasible move. During the transition, savings, insurance, family obligations and access to learning determine which alternatives can be reached. An economy-wide increase in demand is not a bus ticket, a credential or a period of paid retraining. The path connecting the displaced person to the new task is part of the economic problem, even when a long-run account can describe the destination.
The new software makes the task distinction especially visible. In a study published in 2025, Erik Brynjolfsson, Danielle Li and Lindsey Raymond examined an AI assistant introduced among customer-support agents. Across a sample of 5,172 agents, access raised issues resolved per hour by about fifteen percent on average, with larger gains among less experienced and less skilled workers. The system suggested responses to a human agent, who retained discretion over their use. Here the technology helped people perform the task rather than simply eliminating the person between the customer and the answer. This is a dated finding from a particular deployment, not a measure of the employment effect of all generative AI.
The same study leaves the wage question open. The rollout used in the analysis was staggered, and the researchers compared changes in performance while accounting for worker and time differences; it was not a universal experiment on the labor market. They did not observe wages, and the study could not determine the eventual change in the firm's total demand for workers. More issues resolved in an hour might support higher pay, lower prices, greater profit, fewer hours per customer, or some combination. The observed productivity gain does not choose among those outcomes. It makes the division of the gain worth asking about. The assistance that brings a novice closer to an experienced worker may change the value of experience without telling us what either will ultimately be paid.
A task account therefore has to be joined to an account of ownership and bargaining. The machine may belong to the employer, be rented from another business, or be an asset in which workers themselves have a claim. A cheaper product may carry some of the gain to customers who never worked in the industry. More competition for workers with complementary skills may carry another part into wages. Where a displaced worker's alternatives deteriorate, his bargaining position can weaken even if measured productivity at the surviving firm improves. These channels can operate together. Technology changes the feasible arrangements; demand, competition, property and institutions help determine which arrangements are chosen and how their returns are divided.
It would be a mistake to make the worker's security depend on the claim that no task should ever disappear. It would be another to make technological improvement a complete answer to a question about who bears its cost. The first denies the value of producing more with less effort. The second treats an aggregate gain as though it were already a payment to each person who made room for it. A feasible response has to attend to both the creation of useful work and the losses on the way: what skills the new tasks require, how people can acquire them, what supports the interval, and whether workers can claim a part of the productivity they help realize. These questions survive even when the technology works exactly as intended.
A gain in what labor can produce does not settle what a displaced worker can receive.
The next plate gathers the scattered bargains into a fraction. It asks how much of the income produced by an economy reaches labor, and why that fraction can change while productivity rises. The machine supplies one possible route: it can alter tasks, relative costs, the scale of production and the employer's alternative. It does not supply the whole explanation. The worker in this plate is a person encountering a changed possibility; the labor share is an account of millions of such encounters, mixed with changes in firms, prices and ownership. To pass from one to the other requires more than adding up machines. It requires asking where the gains went, and keeping the person who lost the old task visible in the sum.