AI and Jobs: What the Economists Are Actually Saying

In 2025, American companies reported 1.2 million layoffs — a 58% increase from the year before. More than 50,000 were directly attributed to AI. Daron Acemoglu, the MIT economist who won the Nobel Prize in Economics in 2024 (jointly with Simon Johnson and James Robinson), told Fortune that if this trajectory continues, U.S. democracy may not survive the resulting inequality.

That statement sounds alarming. It's also not the full picture of what the data shows — or what Acemoglu's own research framework actually predicts.

Understanding the real debate requires getting past the headlines in both directions.

Acemoglu's Framework

Acemoglu and Restrepo built what they call a task-based model of technological change and employment. The model's central claim is that automation doesn't simply destroy jobs — it changes which tasks within jobs are done by humans versus machines, and those changes have distributional consequences that depend heavily on policy choices, institutional responses, and the pace of new task creation.

The mechanism works in two directions. The displacement effect: automation reduces demand for human labor on the specific tasks machines can now perform. The reinstatement effect: productivity gains from automation create new demand for human labor in tasks that didn't exist before. Whether employment and wages rise or fall depends on which effect dominates.

For most of economic history, new technology created enough new tasks to offset the displacement. The industrial revolution displaced agricultural labor and created factory labor. Computing displaced routine clerical work and created software development, data analysis, and IT support. Each transition was disruptive. Each created more employment than it destroyed.

Acemoglu's concern — and the place where his current work diverges from the conventional economist's comfort — is that AI may change this calculus. LLMs and agent systems can perform cognitive tasks across a much broader range than previous automation, which was mostly limited to structured, repetitive physical or data-processing work. If the technology displaces both routine and non-routine cognitive tasks, the reinstatement effect may not be large enough, fast enough, or distributed well enough to compensate.

His 2024 paper, "The Simple Macroeconomics of AI," models this possibility quantitatively. Under scenarios where AI automates a significant share of cognitive tasks without creating proportional new task categories, the model predicts meaningful long-run declines in labor demand and increases in wage inequality. Under scenarios where AI primarily augments human workers and creates new tasks, the outcomes are positive.

The paper doesn't predict catastrophe. It maps the parameter space and shows that outcomes depend on which path the technology takes — and crucially, on whether policy actively steers toward labor-complementing applications.

What the Current Data Shows

Anthropic published a labor market impacts study analyzing how Claude is being used in real workplaces. Their findings are more granular than most research on this topic because they have access to actual usage data rather than survey responses.

The early pattern: AI assistance is augmenting a wide range of knowledge work, with the highest-use categories being software development, writing, research, and analysis. The augmentation framing holds in most of these categories — humans are doing more with AI assistance, rather than being replaced outright. But the distributional picture within those categories is uneven.

The polarization finding from labor economics keeps appearing in AI-era data as well: automation tends to eliminate middle-skilled jobs — the routine, structured, process-following work that falls between low-skill physical labor and high-skill creative/judgment work. AI is extending this pattern into cognitive work that wasn't previously automatable.

The workers most affected are those who do structured, repeatable cognitive tasks: data entry, routine document processing, standardized customer service, rule-based decision-making. The workers least affected, so far, are those doing work that requires contextual judgment, interpersonal relationships, physical presence, or creative synthesis in unpredictable domains.

The demographics matter here. The workers most concentrated in AI-vulnerable roles tend to be older workers (who have spent careers building expertise in specific procedural domains), women (who are disproportionately represented in administrative and customer service roles), and workers without four-year degrees. If the transition creates new tasks, those tasks tend to require different skills and often different credentials.

The Optimists' Counterargument

The bullish view on AI and employment isn't naive. It has economists behind it too.

David Autor at MIT has documented that labor markets have historically absorbed technological shocks better than predicted because of complementarity — technology tends to make human skills more valuable in adjacent domains. The farmers displaced by mechanized agriculture became the workers in manufacturing and services industries that hadn't existed at agricultural scale. The travel agents displaced by booking platforms became UX researchers, product managers, and customer experience designers at the tech companies that built those platforms.

The argument is that this pattern holds: AI will displace specific tasks, create demand for workers who can work alongside AI, and generate new categories of work that can't currently be anticipated. The adjustment period will be hard for specific workers in specific roles. The long-run equilibrium will have more employment, not less.

The honest version of this view acknowledges two risks: the adjustment period may be longer and more painful than previous technological transitions, and the new jobs may not be accessible to displaced workers without significant retraining. The technological unemployment isn't permanent — but the frictional unemployment during the transition is real.

Where the Policy Debate Actually Is

Acemoglu's prescriptions are specific. He's called for reorienting AI development toward labor-complementing applications rather than labor-replacing ones. He's advocated for wealth taxes to address the concentration of productivity gains in capital rather than labor. He's pushed for active investment in retraining and transition support for displaced workers — not as charity, but as economic policy to enable the reinstatement effect to work.

The mechanism matters here: if AI productivity gains are captured primarily by capital owners and a small class of highly-skilled workers, while the costs of transition are borne by displaced mid-skill workers, you get the inequality scenario without the countervailing new-jobs creation. The technology's outcomes are not predetermined. They're shaped by whether institutions respond to support the transition.

What's largely absent from current policy is proactive design for labor-complementing AI. Most AI research and development is optimizing for automation — for reducing human labor inputs per unit of output. That's economically rational for individual firms. It may not be economically optimal at the aggregate level if it generates unemployment and inequality faster than institutions can compensate.

What This Means in Practice

For businesses deploying AI now, the research suggests a few things worth holding onto.

The most durable human roles aren't "non-automatable" — they're roles where human judgment, relationships, accountability, and contextual adaptation create value that compounds with AI assistance rather than being replaced by it. Organizations that design AI deployment around augmenting those roles tend to retain talent and get better outcomes than those that simply try to reduce headcount.

The displacement is real, but it's not evenly distributed. Strategic thinking about which roles in a business are being affected — and what those workers will do — isn't just an ethical question. It's an operational one. Organizations that manage transitions poorly create disruption that offsets the automation gains.

And the macroeconomic stakes are high enough that the conversation is going to intensify rather than fade. Acemoglu's prediction about democratic stability isn't alarmism — it's grounded in the historical relationship between economic inequality and institutional fragility. The question of how the gains from AI distribute across society is not a technical question. It's a political one, and it's not settled.