IMF | Artificial Intelligence and the Economics of Adjustment

Today’s AI policies will shape tomorrow’s job market.

Artificial intelligence has reignited an old fear—that technology will eliminate work faster than economies can adapt. Variations of this concern appear every time powerful new technologies emerge. What feels different today is the speed, scope, and visibility of AI’s advance, particularly in cognitive tasks long assumed to be uniquely human.

Yet history shows that whenever new technology emerges, economies ultimately undergo deep structural transformation. This allows labor markets to adapt to the potential of the new technology.

Inevitably, many jobs will be changed by AI adoption. Some will be enhanced. Others may be rendered obsolete. But the decline of particular types of jobs is not the same as a sustained reduction in overall employment. If history is any guide, it is not the technology itself that could cause mass unemployment, but the policies deployed in response.

The aggregate impact of AI will depend on how the economy adjusts to it, including whether productivity gains reduce costs, expand demand, and support the creation of new tasks and firms and thereby drive workers and capital toward new uses—enough to offset or outweigh the inevitable loss of some positions or categories of jobs.

The outcomes of the transformation will depend not only on the technological shock itself but more importantly on the policies and institutions that govern the adjustment. If policymakers respond to AI with the wrong set of policies—the kind that delay rather than facilitate adjustment—they risk causing worse rather than better labor market outcomes. Misguided policies could end up slowing growth rather than raising it and increasing inequality rather than containing or reducing it.

Past technological shocks

This time could always be different. But the overwhelming weight of historical evidence supports the claim that innovative technologies do not cause mass unemployment.

The emergence of new general‑purpose technologies is not new. Over the past two centuries, economies have repeatedly absorbed technologies that transformed production, reorganized firms, and displaced entire categories of work—from the steam engine and electrification to computing and the internet.

Each of these technologies disrupted existing jobs and skills. Each provoked anxiety about the future of jobs and work. And each ultimately raised productivity, lowered prices, increased real incomes, and supported higher employment.

Handloom weavers were displaced by mechanized looms. Typists declined as word processing spread. Travel agents were displaced by online booking platforms. In each case, the disappearance of a task was visible and politically salient.

But new jobs emerged—often in sectors that barely existed before. Railways created demand for engineers, mechanics, logistics managers, and entirely new financial services. Electrification enabled industries ranging from household appliances to mass retail and refrigeration. Computing gave rise to software development, data analysis, digital design, and a wide range of professional services. Aggregate employment did not collapse. It ultimately increased.

It is important to recognize that these gains were neither automatic nor immediate. They were staggered and gradual and required diffusion, complementary investment, and institutional adaptation. Nonetheless, a systematic review of the economic literature published in 2023, covering more than 100 studies, showed that the labor displacing effect of technology is more than offset by labor creation.

What changed was not simply the number of jobs, but the content of work. As recent IMF research shows, jobs evolve not only because workers move across occupations, but because of shifts in the skills required within the same occupation. Roughly 1 in 10 job vacancies in advanced economies now lists a new skill. This is how technological change usually unfolds: not through wholesale job destruction but by reshaping what workers do and the skills they need.

Markets’ absorption of new technologies reflects a well‑known economic dynamic. As processes become more efficient, costs fall. Lower costs translate into lower prices. Lower prices stimulate demand. And higher demand supports higher output and employment. Economists often refer to this as the Jevons paradox, and it has been observed repeatedly across sectors, from energy to transport to information processing.

The key point is that new technology expands economic activity—by opening new markets, increasing scale, and supporting entirely new forms of production, consumption, and employment.

What this means for AI

This is likely to be the net impact of AI. By reducing the cost of analysis, prediction, communication, coordination, and—increasingly in its agentic form—action, AI makes a wide range of services cheaper and more scalable. This will eventually increase demand, while also enabling new products, services, and firms.

A recurring mistake in today’s debate is a modern version of what economists have long called the “lump‑of‑labor fallacy”—the belief that there is a fixed amount of work to be done, so that if machines do more, there will be less for people to do. History tells a different story: Technological progress has created more jobs than it has destroyed. And as IMF research shows, when AI complements human labor and productivity gains are sufficiently large, AI adoption can lead to higher growth and incomes for most workers.

None of this implies that the adjustment will be smooth or costless. Even when technological change ultimately supports net job creation, the transition can be disruptive. Three issues in particular deserve close attention.

First, we do not yet know what the new work will look like, who will do it, or where it will be located. And the social and political disruption created by that uncertainty, at both the individual and societal levels, can be significant. IMF staff estimates find that approximately 40 percent of jobs globally could be affected by AI in some way—not necessarily eliminated, but changed. That includes changes in task composition, skill requirements, and organizational structure. Many jobs will be enhanced by AI. Some may be rendered obsolete. The structure of labor demand could also change. Emerging evidence suggests near term gains may be strongest for high- and low-skill workers, while demand for middle-skill and entry-level positions could weaken. However, predicting the long-term trajectory of trends in labor demand is exceptionally uncertain. Whatever shifts do materialize will have important political economy consequences that policymakers will need to manage.

Second, AI could accelerate churn in labor markets that may present challenges for some of those markets. The scale of preexisting labor market churn, moreover, is often underappreciated. In the United States, total nonfarm employment is roughly 160 million, and there are approximately 60 million hires and 60 million separations every year. This extraordinary scale of job creation and destruction is happening every day throughout every corner of society. Countries with less flexible labor markets may struggle to reallocate resources in the world of AI more than those with a higher degree of churn.

Third, the labor market adjustment caused by AI could be slowed or distorted by policy choices, institutional frictions, or market failures. We have seen this before. Early uses of electricity focused on powering existing factory layouts rather than reorganizing production lines to seize the full potential of the new technology. Early uses of computing automated clerical tasks before enabling entirely new businesses and organizational forms. In both cases, the largest productivity gains came later, once complementary investments and institutional changes caught up.

Economic historians sometimes describe this dynamic as an “Engels’ pause,” after Friedrich Engels’ analysis of the combination of rapid economic growth and stagnant wage growth Britain experienced in the first half of the 19th century. The term has come to denote a period in which new technologies diffuse through the economy, disrupting existing structures, before new business models and activities fully emerge. During that period, gains can appear uneven, and labor market adjustment can be painful. Distributional changes may be material and cause social and political disruption.

Policies to facilitate adjustment

The task for policymakers will be to maximize the potential benefits from AI while insuring against the potential negative consequences. This won’t be easy: Realizing the benefits from AI will require large shifts in labor and capital across the economy, which can be disruptive if the process of job reallocation is prolonged or poorly managed. What should policy aim to do and not do?

The overarching response to a structural shock like AI should be structural policies—designed to facilitate adjustment rather than prevent it. These include labor market policies that support mobility and reemployment, product market policies that promote competition, and financial and legal frameworks that allow capital and assets to be reallocated efficiently and productively.

Labor market policies are particularly important. Many existing labor market institutions are designed to deal with cyclical, not structural, shocks. Furlough programs, job retention subsidies, and temporary layoff protections can be highly effective when aggregate demand falls temporarily and then recovers. Such measures are much less effective when entire sectors need to shrink, and new ones need to expand. What works for a recession does not necessarily work for a technological transition.

Policies that help workers navigate transitions without locking economies into outdated structures can also play a useful role, especially retraining and upskilling programs. These programs should be widely accessible and designed to maintain private sector incentives for both businesses and employees so that new employment relationships can take hold without government support.

But policymakers should not rely too much on these policies, since the track record of many active labor market policies is mixed and may not sufficiently address the scale of the challenge posed by AI. Governments often lack the necessary information, incentives, and institutional capacity, particularly in fast‑moving technological environments. AI itself will provide a significant opportunity both for upgrading active labor market policies and for turbocharging the employment services industry, given its capacity for personalized education and reducing information frictions.

On the other hand, policies that slow adjustment, by protecting specific jobs, firms, or sectors, would delay reallocation, reduce productivity growth, and ultimately lead to worse labor market outcomes. There is an understandable tendency to respond to disruption with protection. But this can end up harming the very people it is intended to help. If policymakers make it more expensive for companies to employ or fire workers, companies will pay less or not create jobs at all.

AI regulation

A similar nuanced approach should be applied to AI regulation.

Guardrails are clearly necessary in some areas, including when it comes to cybersecurity and the protection of children. In these cases, risks are concrete and externalities are clear. But there should be a presumption against protective action unless there is strong evidence of harm or the presence of clear risks. A rush toward regulation without a clear rationale—for example, generalized fear about job losses—could leave society worse off because of prolonged misallocation of resources and drifting away from the technological frontier.

AI regulation that focuses on permission structures rather than restrictions on AI use can be productive. For example, heavily regulated sectors like health care and finance may need additional regulatory certainty on the suitability of using AI systems in order to realize the productivity benefits of AI adoption.

More broadly, regulation should promote business dynamism. This entails reducing barriers to entry to avoid regulatory capture and excessive market concentration and maintain the significant competitive dynamics currently evident in the AI ecosystem. Policymakers also need to ensure that bankruptcy and restructuring frameworks are working efficiently to support the speedy reallocation of resources.

Developing economies

The stakes are particularly high for emerging markets and low‑income countries.

For these economies, AI presents a genuine leapfrogging opportunity. Digital delivery of services can overcome physical infrastructure constraints. AI‑enabled diagnostics can expand health care access, and automated compliance tools can lower the cost of formality for small firms. Governments can also use AI to improve tax administration, customs, and social protection delivery.

But AI-related risks specific to emerging markets and low‑income countries are substantial. If AI leads to sustained productivity gains in advanced economies first, income gaps could grow. And these gaps would widen further if economic rents from AI became geographically concentrated. Capital could flow uphill, diverting much‑needed financing away from lower‑income countries.

These risks are amplified by existing structural constraints in some developing economies: slow labor reallocation, barriers to firms’ entry and exit, limited access to financing, weak legal systems, and ill-defined property rights.

Policies should therefore be tailored to countries’ level of preparedness. More developed and better-prepared economies should focus on innovation and diffusion—through R&D investment, improved access to financing, and a business environment that fosters innovative firms—along with regulatory frameworks that enable safe and widespread AI use. Lower-income and less-prepared economies should focus initially on building digital infrastructure—especially reliable and affordable power generation—and on education, with a greater focus on earlier attachment to the labor market, lifelong learning, and skills that complement rather than compete with technology. These investments can support AI adoption while advancing broader development goals; they should be embedded in a strategy that safeguards fiscal sustainability and is aligned with absorptive capacity.

Role of the IMF

History offers many examples of policies designed to slow structural change that ended up entrenching inefficiency, delaying recovery, and worsening outcomes. If protection becomes the dominant response, the economic and social disruptions associated with AI adoption could overwhelm the potential benefits. A prolonged and politically difficult period of weak growth and stalled adjustment could ensue—another version of Engels’ pause.

In the worst case, a protective response could cause a political backlash against the technological progress and creative destruction that underpins long-term improvements in living standards. Engels’ pause in the United Kingdom coincided with the rise of the Luddite movement. And while Engels did not coin the term Engels’ pause, his experiences during this period informed his subsequent collaboration with Karl Marx on the development of the political philosophy of Marxism. Though neither the Luddites nor Marxists succeeded in 19th century Britain as a response to Engels’ pause, variants elsewhere have damaged standards of living in the subsequent two centuries.

The goal of policy should not be to protect specific jobs, companies, or industries. It should be to incentivize employees and businesses to adapt and to unlock the productivity gains from AI. This requires flexibility, dynamism and sound structural policies—not stasis.

Institutions like the IMF can help realize this crucial policy goal. We are strengthening our surveillance of AI‑related structural changes. We are supporting members in designing reform strategies and thinking through trade-offs. And we are facilitating international cooperation and knowledge sharing on AI‑related best practices to help avoid some of the pitfalls of earlier periods of economic transformation.

We also see AI as a major opportunity to enhance our own operations. Used effectively, AI can help us do more with the same resources, improve the quality and speed of our analysis, and further increase the value we deliver to our membership.

We do not know what the future of work will look like. But it will be heavily shaped by the policies and institutions that govern how economies adapt to technological change. Our choices today will determine whether this transformation lifts growth and prosperity—or leaves our societies, politics, and institutions struggling to catch up.

 

 

Compliments of the International Monetary Fund