Artificial intelligence seems to be growing more powerful with each passing day. A model recently released by OpenAI can earn top scores on law school final exams, and ChatGPT and Gemini can outperform many young college graduates on tasks in their jobs. The White House’s “AI Action Plan” aims, if anything, to speed up the development and use of AI in the workplace. This all leaves many fearful that, within a few short years, AI will gobble up most work now performed by college graduates or even those with professional degrees, while AI-powered robots, drones, and autonomous vehicles take many blue-collar jobs as well.

But we’ve been here before. About 20 years ago, we were told that tens of millions of U.S. jobs would be outsourced to China or other countries—which didn’t happen. In the late 1950s and early ’60s, America experienced “automation anxiety”—when people learned about the power of computers to do many tasks—and concerns about high unemployment were widespread. Time wrote of “the automation jobless” and the fear at the highest levels of government that new jobs wouldn’t be created. And, of course, the Luddites in Britain more than two centuries ago smashed industrial machinery that they believed would cause massive joblessness, after some had lost their jobs to automation.

But these fears have mostly proven unfounded. About 250 years after the beginning of the Industrial Revolution, America and most other industrial countries are typically awash in jobs: Though U.S. labor markets have recently softened, vacant jobs remain roughly as high in number as unemployed workers. If anything, the United States has experienced more worker shortages in recent years than unemployment.

At the same time, millions of workers can be hurt by automation, either through lost work or lower wages and benefits. And AI has the potential to have much more significant labor market impacts than past modes of automation, as it can perform tasks now done by large percentages of current workers and as it becomes more powerful over time.

So how should we think about the problem of AI and jobs, and what should we do about it? While AI will no doubt destroy jobs, it will also create many new ones. I focus below on these new jobs and the skills that workers will need to thrive in them. As AI continues to improve, workers and their employers will need to continually adapt, while education practices and public policies must be put in place to help them make these adjustments.

The Economics of Automation and the Job Market

There is no question that machines like calculators, computers, and robots replace some workers as they perform tasks that these workers had done in the past. The negative economic impacts of this are well known: They include worker displacement and lower wages for some who remain employed. But many don’t realize that automation has other, more positive economic effects on workers, as well as on consumers and businesses.

Basically, automation raises worker productivity and lowers the costs of producing goods and services. By reducing costs, it should also reduce prices—which raises our incomes and wealth in real (inflation-adjusted) terms, making us all more prosperous. Since consumers are better off, they spend more than before. If automation reduces specific prices substantially—think of Henry Ford’s first assembly lines in 1913-14, personal computers in the 1980s, or smartphones more recently—consumer demand for the product in question will increase dramatically and employment in the automating industry will rise.

For instance, employment in the automobile industry rose from a very modest level in the 1910s to hundreds of thousands by 1929. In more recent decades, more than one million U.S. workers have produced automobiles and parts (despite the automation and globalization of auto production) while another 1.3 million work in auto dealerships. More recently, Apple’s U.S. workforce has risen from 5,000 in 1998 to 80,000 today, while related employment in everything from app creation to technical services could be as high as two million. In addition, richer consumers spend more on many other goods and services, leading to more jobs elsewhere too.

But some workers are, in fact, hurt by automation. Most clearly, some are directly displaced—and they might not be the ones hired into the newly created jobs. In fact, some entire job categories may be eliminated by automation, though new ones are formed in their place. For instance, Henry Ford’s assembly line ended work for the skilled craftsmen who produced cars before then and it left them worse off, even as hundreds of thousands of other workers got well-compensated jobs on assembly lines. Displaced workers can suffer large and lasting losses, sometimes not regaining employment.

More broadly, entire classes of workers can be hurt by certain kinds of technical change. For instance, the digital revolution after 1980 hurt most non-college workers, especially those who were employed on factory assembly lines or as clerical workers in offices. Facing less demand in their old jobs, they began competing with similar workers in other occupations and industries, a process that drove down all of their wages. In contrast, those with college and especially professional degrees saw more demand for their labor, as personal computers made it easier to perform and disseminate their work.

Economists think that automation substitutes for some workers and complements others, leaving the first group worse off and the second better off. In general, technical change can be skill-biased: It tends to hurt the less educated and help the more educated, thereby raising inequality. To be clear, other developments of the past several decades—like declining unionism and rising imports from China—have also led to higher inequality between the college- and non-college-educated, but technical change was no doubt a major force.

But this is not the end of the story. If inequality rises between those with college degrees and those without them, more people will have an incentive to get college degrees. As the supply of college grads rises, their earnings will drop, and gaps between the two groups will shrink. Harvard economists Claudia Goldin and Lawrence Katz call this the “race between education and technology, ” with higher education levels offsetting the higher inequality that the new technologies initially create. In this process, workers have incentives to turn themselves into complements to automation, becoming better off and sharing in the fruits of automation-led prosperity.

Note that government policy plays an important role in this process. It can directly help those who lose their jobs with replacement income through programs like unemployment insurance (UI) while providing more of the education or training that will complement the new automation. We subsidize those who work at lower wages through the earned income tax credit, and those with children through the child tax credit. The college-educated and others who benefit from automation—like the owners of the machines (or capital)—face higher income taxes to finance the new public spending needed for education or other worker supports.

But when new general forms of automation appear, like AI, the negative effects could be greater or more widespread than before. In that case, we might need newer and more innovative education practices, along with new sources of funding to help workers adjust. And we must use antitrust and other reforms to keep markets competitive so that consumers benefit from lower prices and workers benefit from new jobs.

In sum, automation can have both positive and negative effects on workers. Many will need to adjust their skill sets in order to benefit more from these changes, and new policies must be implemented to help them do so.

AI: Is This Time Different?

While all of these economic forces might be at play, there are some characteristics of AI that make it feel different from the assembly line or mechanized looms. The range of what it can do now and will be able to do in the future is quite astounding. Generative AI (or GAI) can not only write computer code and first drafts of essays; it can also generate pictures, poems, and music. Indeed, it can absorb much of what is available online today and use it to generate these products, though there are already important questions about the quality, creativity, and emotional “humanness” of its creations.

Over time, AI will likely keep getting better and better. “Agentic” AI will soon be able to act autonomously, taking on complex tasks and making decisions based on benefits and costs. (Of course, human workers will still be needed to monitor these decisions and the accuracy of the inferences AI makes.)

All of this means that AI will substantially increase our productivity over time, thus raising our incomes and standards of living. But these increases in well-being will likely occur unevenly. AI is perhaps already powerful enough to reduce employer demand for entry-level college graduates, whose unemployment rate has been unusually high in 2025 at 4.6 percent (though this likely reflects other factors too, like general labor market softness and employer uncertainty). In a set of occupations with fairly high exposure to AI, such as software development and customer service, the rise in unemployment among young workers is notable.

An interesting possibility is that while college graduates could face more replacement by AI, workers without college degrees might now become more productive. For instance, AI might enable mechanics to better identify the problems in some machinery or nursing assistants to better assess a patient’s condition and needs, leading to higher earnings for them. If that were to happen, AI could help reduce the inequality that now exists between college grads and other workers.

On the other hand, while college grads might have more negative exposure to AI, their ability to adapt to the new reality of AI will likely be better. College graduates might more quickly perceive the threats to their jobs that AI poses and proactively respond by learning to perform new tasks using AI. Employers, who currently invest much more in training college graduates than other workers, might also be more willing to retrain their college graduates.

The fact that AI will continue to improve and outperform humans on more and more tasks will make it harder for workers to adjust, since a set of skills that might complement AI today could be replaced by it tomorrow. Workers might feel like they are on a treadmill where they can never stop obtaining new skills or knowledge. For example, those who invested in training to become computer programmers just a few years ago have started to face obsolescence as AI develops these capabilities—though perhaps not as rapidly as some claim. In the “race between education and technology,” it may become harder than before for educated workers to catch up and stay even with the machines.

At the same time, there are human qualities that AI will almost certainly not match anytime soon. We will remain better at making judgments based on both data and a range of subjective factors. Empathy, creativity, and skill at handling a variety of social interactions will all remain the realm of humans for a long time. And the jobs that heavily depend on these qualities—including care jobs, certainly, but also those where creativity still matters for job performance, and even medical, legal, or financial jobs where careful judgments must precede action—are areas where humans will long retain an edge.

So How Can Policy Help Workers?

Government can and must play an important role in guiding: a) how software developers make AI, b) how employers implement it in the workplace, and c) how workers adapt. The goal should be to minimize the displacement costs of the new technology and make sure its benefits are widely shared.

Regarding AI software development, federal grants to developers can reward efforts to make AI more “human-centered”—in other words, complementary to a range of worker skills reflecting human qualities that AI will not successfully mimic. In the absence of such efforts, AI developers might focus on replacing workers; in their presence, AI could be developed in ways that more carefully augment workers’ skills and allow them a greater role in its use.

Market forces alone will mostly drive AI developers to appeal to private employers, whose only concern might be minimizing their labor costs and thus raising profits. To soften this tendency, we can use public dollars to develop AI to be more complementary to (or usable by) a wide range of workers, especially those without college degrees. We can also use regulations to influence AI development—though with the major caveat that excessive regulation could adversely affect our ability to develop and implement AI, thereby potentially depriving us of its enormous productivity gains.

Government can also influence employers’ decisions on how to implement AI. When AI takes over some of the tasks previously performed by workers, employers will have a lot of discretion over which employees they retrain to perform new tasks and which they let go. As noted, these choices can easily be influenced by employer biases over employee education and background, not to mention race and gender. Meanwhile, the tax system can lead to a broader bias in favor of firing workers, as the federal tax code rewards employers who replace labor with new equipment through generous depreciation write-offs.

Since worker displacement imposes very large costs on workers and society (see many Rust Belt communities over the last few decades as manufacturing jobs disappeared), and since employer biases can play a major role in generating these costs, it makes sense for government to strengthen employers’ incentives to retain and retrain their employees. For instance, we could think about imposing a modest “displacement tax” on employers, in which they pay a fee for every worker displaced. In addition, we could perhaps consider subsidizing retraining, especially for workers without a college education, who are less likely to be retained and retrained otherwise.

A “displacement tax” would not be a tax on AI, but only on how it is used. It should not be so high as to discourage AI adoption or the hiring of workers in the first place, but just high enough to influence how employers treat workers when they implement AI. There is already some precedent for this: The taxes on employers to finance UI are “experience rated,” meaning they are higher for employers who generate more layoffs than for those who generate fewer. But worker displacement from permanent layoffs—the kind that are likely to occur as AI transforms the economy—is much more socially costly than displacement from temporary layoffs. Reducing unnecessary permanent layoffs could therefore be seen as a “public good” that employers will not provide on their own without public incentives. The proceeds of the tax could then be used to subsidize worker retraining. We could also just try to soften the tax code biases that now favor automation, perhaps by making tax write-offs for depreciation of new equipment a bit less generous.

Finally, governments at all levels can improve education and training options to make workers more complementary with AI. In K-12, career and technical, and higher education, we can promote AI literacy and, more broadly, the human skills that complement AI, like judgment, critical thinking, and social/communication abilities. We can encourage colleges to provide more experiential learning for students, and firms to adopt more work-based learning options (like apprenticeships) to provide the entry-level experience that AI might be eliminating.

As workers face potential or actual displacement by AI, we can also improve their ability to adjust in a variety of ways. For instance, we can promote payroll deductions into “lifelong learning” (or “skill savings”) accounts, building on a number of pilot programs across the country. These can be mandatory or voluntary. Contributions could be deducted from taxable earnings to encourage greater use of these accounts, and each dollar invested could be matched with public funds, particularly for lower-wage workers whose balances will otherwise grow much more slowly than those of higher-wage workers at a given deduction rate.

The federal government could also create an “Automation Adjustment Assistance” program, modeled after Trade Adjustment Assistance (TAA). For decades, the latter gave displaced workers funds for training and income support above what they received from UI, if they could certify that their displacement was caused by imports. We could do the same with AI-based automation. Though TAA has long been considered relatively ineffective, recent research suggests it became quite effective over time, especially when workers were given more guidance and training for in-demand and well-paid jobs. (Unfortunately, TAA was not reauthorized after 2022, meaning workers who were already deemed eligible can continue to receive benefits but no new petitions are being considered.)

More broadly, we need workforce training programs to be nimble and to prepare workers for high-demand jobs (as measured by the latest data), even when those jobs are rapidly evolving. Our community colleges and other training providers must partner with their regional employers on a continuing basis to ensure that they are helping to train workers for jobs that are hard to fill and that pay workers well. They must be able to respond quickly when automation changes skill requirements—which tends to be more easily done in non-credit than for-credit programs, as the former are less subject to bureaucratic governance.

This, in turn, would require that we find new ways to finance enrolled students. Until recently, lower-income students could get Pell grants or federal loans only to finance for-credit programs of a certain minimum length. But because of the One Big Beautiful Bill Act passed last summer, Pell grants are now available for shorter programs and even some that are not for credit, as long as they meet a set of criteria (such as providing average earnings above those of local high-school graduates). Alternatively, we might use other models of financing students, such as income share agreements or outcome-based loans. In these newer models, which have already been implemented in a number of specific cases by private organizations and state and local governments, students receive loans for higher education or training and only repay when their earnings have reached certain minimum thresholds (usually around $40,000 per year).

We should also remember that AI will itself improve the quality of education and training provided virtually and make such education more accessible. Rather than going to sit in community college classrooms, workers will be better able to obtain retraining in their workplaces or at home. “Intelligent tutoring systems” will help them interact with online trainers and provide them with individually tailored assistance. The costs of providing such training will hopefully decline over time, making it accessible to larger numbers of students.

Federal and state governments have a number of additional policy options for helping workers who are laid off or lose income due to AI. They could strengthen both the funding of and services provided by UI. They could provide “wage insurance” to bolster the lower wages many workers receive in new jobs after losing earlier ones. The child tax credit and earned income tax credit could be strengthened as well, so that the tax code provides more help to families with children and low earnings in the era of AI. And, as AI will probably raise employer profits, government could encourage more profit-sharing with workers—for instance, by making the tax benefits of employee stock ownership plans (ESOPs) more generous than they are now. Giving workers some voice in how AI is implemented in the workplace—whether through unions or some other platform—would also help ensure that its benefits are widely shared.

On the other hand, I am very reluctant to embrace universal basic income (UBI) programs, as some have advocated as a response to AI. UBI is often promoted by those who expect massive job disappearance due to AI, which is a prediction I do not find credible. UBI would encourage workers to forego low-wage jobs that might be available and would swell our already enormous federal deficits. For those who are displaced and cannot find alternative work, publicly subsidized jobs are a more sensible way to go. Under this approach, public funds could subsidize either public- or private-sector jobs, which disadvantaged workers or those with lengthy spells of unemployment could hold for limited durations.

Toward Widely Shared Prosperity in an AI-Driven Future

All of these newer efforts will require new research and evaluation, based on new data sources, to measure the costs and benefits of AI implementation and policy adjustments, and to whom they accrue. A willingness to update our policies and programs based on such evidence will be critical to our success in helping workers adjust to the new labor market they will face.

Overall, we should take a clear-eyed approach to the rise of AI that will help us enjoy the technology’s enormous benefits while relieving the disruption and pain that will likely occur as AI is implemented. AI will dramatically increase worker productivity in many realms, which should raise earnings among those who can use it. This might be particularly beneficial to many workers without college degrees who today lack the reading and writing skills needed in many higher-paying jobs. At the same time, millions of workers will likely see AI take over the tasks they currently perform at work, and they will need to develop new skill sets, particularly for the dozens of new job categories that will appear in response to AI. We must implement a range of public policies to help them do so. If those policies are effective, then we can make sure that the income and wealth benefits of AI are widely shared.