Headlines ask whether robots will "take" a job, as if a job were a single thing that a machine either can or cannot do. Since the early 2000s, economists have found it more useful to think about tasks. In the task approach, a job is a bundle of tasks, and technology usually takes over some tasks while leaving others, or creating new ones.

Routine and nonroutine

In a 2003 paper, David Autor, Frank Levy and Richard Murnane sorted tasks by two questions. Is the task cognitive or manual? And is it routine? A task is routine when it follows explicit rules that can be written down step by step. Keeping a ledger, sorting mail and running a machine on an assembly line are routine. Computers are very good at these because a programmer can spell out every step.

Nonroutine tasks come in two kinds. Nonroutine cognitive tasks involve judgment, persuasion, creativity or solving new problems, like managing a team or diagnosing an unusual illness. Nonroutine manual tasks involve dealing with a messy physical world and other people, like cleaning a hotel room, cutting hair or caring for a frail patient. For decades, computers did poorly at both.

Two paradoxes

Why should folding laundry be harder for a machine than calculating a payroll? The philosopher Michael Polanyi observed in the 1960s that people know far more than they can put into words. You can recognize a friend's face in a crowd, but you could not write down the rules you used. Autor named this Polanyi's paradox: Polanyi's paradox says that we know more than we can tell, so many skills resist being written as rules. Traditional programming could only automate what someone could explain.

Robotics researchers noticed a related puzzle. In the 1980s Hans Moravec and others pointed out that computers could beat experts at chess long before they could walk across a cluttered room. Moravec's paradox is that skills hard for people, like advanced math, can be easy for machines, while easy human skills like walking and grasping are hard. Abilities that evolution polished over millions of years feel effortless to us precisely because they are so deeply built in.

The hollowing of the middle

Put these ideas together and you get a prediction that turned out to fit the data. From the 1980s through the 2000s, many routine middle-wage jobs shrank: bookkeepers, clerks, typists, bank back-office staff and many factory workers. Jobs at the top, full of nonroutine cognitive tasks, grew. So did many low-wage service jobs full of nonroutine manual tasks, from food service to home health care. Economists call this job polarization: growth at the high-paid and low-paid ends while middle-wage routine jobs shrink. It appeared in the United States and much of Western Europe, though its size varied.

Displacement and new work

The task approach also helps explain why automation has not ended work. Daron Acemoglu and Pascual Restrepo describe several forces acting at once. The displacement effect is the direct loss when machines take over tasks people did. The productivity effect is the extra demand for labor elsewhere when goods get cheaper and incomes rise. And the reinstatement effect is the creation of new tasks in which people have the advantage, such as software design, social media management or ultrasound technology. Whether a technology helps or hurts workers overall depends on the balance among these forces.

New work is not a small side effect. A team led by Autor matched job titles in census records back to 1940 and found that about 60 percent of American workers in 2018 held job titles that did not exist in 1940. They also found that new work tended to appear where technology complemented people, while automation hit where technology substituted for them, and that new work has lately been concentrated in high-paid professional jobs and low-paid personal services.

Worked example: will AI replace radiologists?

In 2016 the AI pioneer Geoffrey Hinton remarked that hospitals should stop training radiologists, because within five years or so image-reading software would outperform them. Image recognition did improve dramatically. Yet the number of radiologists did not fall; in the United States demand for them kept rising and shortages were widely reported. The task approach shows why. Reading scans is a large part of a radiologist's day, but the job also includes consulting with other doctors, choosing which images to take, performing guided procedures, explaining results to patients, catching errors in software and bearing legal responsibility. Radiologists did not vanish because AI took over some tasks while the job contained many others, and faster scans also raised the number of scans ordered. This does not prove radiologists are safe forever. It shows that the right question is not "Can AI do this job?" but "Which tasks within this job can AI do, what happens to demand when those tasks get cheaper, and what new tasks appear?"