The Jobs AI Takes and the People It Makes

The first question

From the first public encounters with generative AI, one question dominated: how many jobs will it take?

The concern was understandable. A machine that could write, draw, summarize, analyze, and program appeared to reach directly into work once considered safely human. Earlier automation had transformed factories and farms. This time, lawyers, designers, teachers, accountants, writers, and software engineers could imagine themselves inside the machine's reach.

Some companies encouraged the anxiety. AI was presented not only as a new tool but as a way to do more with fewer people. Headlines counted occupations at risk. Demonstrations showed a model completing in seconds what a person might bill hours to produce. The implied future was one in which human labor became increasingly unnecessary.

That future is possible in parts. Tasks will be automated. Some positions will vanish, and some organizations will use AI primarily to reduce headcount. But a forecast based only on substitution assumes that the economy contains a fixed amount of useful work. History suggests otherwise.

Jobs are bundles of tasks

A job is rarely one activity. It is a bundle of tasks, relationships, judgment, knowledge, responsibility, and routine. Technology can remove one part of that bundle while making another part more important.

The International Labour Organization estimates that one in four jobs worldwide has some exposure to generative AI. Yet its central conclusion is not that one in four jobs will disappear. Because most occupations still require human input across many tasks, transformation is more likely than complete replacement. ^[1]

This distinction matters. Exposure measures what AI may be able to do. It does not tell us whether a company will adopt it, whether customers will accept it, whether the output will be reliable, or whether the remaining human work will become more or less valuable.

A doctor who uses AI to prepare notes is still a doctor. A programmer who asks a model to produce routine code still has to decide what should be built, understand the system around it, evaluate the result, and remain accountable when it fails. The occupation survives, but its center of gravity moves.

Creation is harder to see

Job destruction is visible. A company announces layoffs. A department closes. A familiar occupation declines. Job creation is harder to recognize because the new work often lacks a name until it already exists.

Research by David Autor and his colleagues found that the majority of U.S. employment in 2018 was in job specialties introduced after 1940. New work counterbalanced much of the labor demand lost to task-displacing automation, although its rewards were uneven and increasingly polarized. ^[2]

That history does not prove AI will create more jobs than it removes. AI differs from previous technologies in speed, reach, and its ability to operate across many cognitive tasks. It does show why counting only today's occupations produces a distorted picture. The future contains work we cannot yet name because the products, expectations, and problems around it do not yet exist.

The World Economic Forum projects substantial labor market churn by 2030, with 170 million roles created and 92 million displaced across the major technological, demographic, environmental, and economic forces it studied. That is not a forecast for AI alone, and it comes from employer expectations rather than certainty. Still, it reinforces a useful point: creation and displacement can happen at the same time. ^[3]

Becoming possible

The employment debate usually treats people as occupants of jobs. It asks whether a position survives and whether a worker can move into another one. But AI may change something more personal: what an individual believes they are capable of doing.

A person who cannot program can describe a product and produce working software. Someone without a design team can develop a visual identity. A small business can analyze information that once required a specialist. A writer working in a second language can communicate with greater confidence. A researcher can begin exploring an adjacent discipline without spending months learning its vocabulary first.

None of these people instantly becomes an expert. AI can generate plausible mistakes, conceal gaps in understanding, and make an amateur feel competent before they are competent. But the first barrier changes. The question moves from “Do I possess every technical skill required to begin?” to “Can I define the problem, direct the tools, judge the result, and keep learning?”

This is not merely efficiency. It is agency. People can attempt projects that would previously have required more money, more credentials, more employees, or permission from an institution. Some of those attempts will become businesses and professions. Others will remain personal achievements. Both expand the territory of human possibility.

Only scratching the surface

Today, much of generative AI use remains concentrated in practical guidance, information seeking, writing, and coding. We ask it to improve messages, summarize meetings, explain unfamiliar subjects, and accelerate work we were already doing. ^[4]

That is the connection to the better-emails problem. If we are still using a potentially general technology mainly to make existing knowledge work faster, then we should be cautious about treating the current labor market as its final destination.

The first stage of a technology often imitates what came before it. Early websites resembled brochures. Early television borrowed from radio and theater. Early smartphones carried familiar desktop functions into a smaller screen before entirely new services emerged around location, cameras, sensors, and permanent connectivity.

AI is still largely being inserted into existing workflows. Its more consequential employment effects may arrive when people redesign the workflow, the organization, or the product around capabilities that were previously unavailable. That is when new work becomes imaginable.

Optimism needs honesty

Believing that AI will create opportunity does not mean every displaced worker will benefit from it.

New jobs may appear in different places, require different skills, or arrive years after older jobs disappear. A fifty-year-old administrator cannot pay a mortgage with the promise that a new profession will eventually be created. A copywriter who loses steady work is not made whole because a successful founder can now build a company with fewer employees.

Historical transitions have created wealth while imposing concentrated costs. Research on the automation of telephone switching found that overall employment recovered for later generations, but the operators directly affected were more likely to move into lower-paid work or leave employment. ^[5]

This is the difference between being right in aggregate and being just in practice. Even if AI creates more work than it destroys, the transition can widen inequality and weaken communities unless people have realistic paths into the new opportunities.

Optimism is credible only when it makes room for that pain.

The choices inside the outcome

Whether AI augments people or removes them is not determined by the model alone. It is shaped by product design, management decisions, education, labor markets, and public policy.

A company can use AI to remove entry-level positions, or it can give entry-level employees greater reach and responsibility. A school can treat AI only as a cheating problem, or teach students how to question, verify, and build with it. A government can wait for disruption, or invest in portable benefits, transition support, lifelong learning, and access to the tools that create new work.

Businesses also have a reason to look beyond headcount reduction. Replacing labor captures a finite saving. Expanding what an organization can offer may create entirely new demand. The largest gains may come not from producing the same output with fewer people, but from serving customers and solving problems that were previously uneconomic.

The goal should not be to preserve every existing task. It should be to make the gains from new capability broad enough that people can move toward better work instead of simply competing with cheaper machines.

More than a job count

The future of work cannot be reduced to whether the total number of jobs rises or falls. A society can achieve net job creation while leaving millions of people behind. It can also preserve employment while degrading autonomy, wages, and meaning.

The better question is what kind of capability AI distributes, to whom, and under whose control.

I am in the job-creation camp, but not because technology automatically produces a happy ending. I believe people continually find new things worth doing when the cost of acting on an idea falls. AI lowers that cost across an unusual range of human activity. It gives an individual access to forms of analysis, expression, and production that once belonged mainly to specialists and organizations.

We are seeing this while the technology is still scratching the surface. Most people are not yet reorganizing their lives or creating entirely new professions around AI. They are learning what it can do by improving an email, generating an image, analyzing a document, or building a small piece of software.

Those modest uses are not the end of the story. They are how people begin to discover that the boundary around what they can attempt has moved.

AI will take jobs. It will also create jobs. Its deepest effect may be harder to count: people becoming builders, creators, researchers, and entrepreneurs before they would ever have used those words to describe themselves.