Will AI take your job? The question has produced a flood of forecasts, and they point in different directions because they measure different things. Some count the tasks that software could theoretically perform, which yields alarming percentages. Others count actual hiring and employment, which so far show something much smaller and more uneven. This article separates what has been measured from what has been predicted, using studies from the International Monetary Fund, Stanford and Yale, and flags where the evidence is early or disputed.

What is the difference between a job being “exposed” to AI and a job being lost?

Most of the scary numbers measure exposure. An occupation is exposed if AI could do, or speed up, a meaningful share of its tasks. Exposure says nothing about whether employers will actually use the technology, whether it works reliably enough, or whether the savings will cut headcount or just raise output.

It also cuts both ways. A highly exposed job might be partly automated, with the worker shifting to other tasks. Or the technology might make each worker so much more productive that demand for the work grows. Economists call the first effect substitution and the second complementarity, and the studies below disagree mostly on which one dominates.

Keeping that split in mind makes most headlines easier to read. “40 percent of jobs exposed” and “40 percent of jobs gone” are different claims, and only the first has any support.

How many jobs are exposed to AI, according to the IMF?

In a January 2024 analysis, International Monetary Fund staff estimated that almost 40 percent of global employment is exposed to AI. The share is higher in wealthy economies, about 60 percent, because they have more cognitive, office-based work. For emerging markets the estimate is about 40 percent, and for low-income countries about 26 percent.

The IMF’s reading was mixed. In advanced economies, roughly half of the exposed jobs might benefit, as AI complements the worker and raises productivity. For the other half, AI could take over key tasks, which could reduce labor demand, wages and hiring. The analysis also concluded that AI would likely worsen overall inequality in most scenarios, partly because gains tend to flow to higher earners.

This is a model-based projection built on assumptions about which tasks are automatable. It is a map of where change could land, not a count of lost positions.

What does the data on actual employment show?

On aggregate numbers, the effect has been hard to find. The Budget Lab at Yale published an analysis on October 1, 2025, measuring how much the mix of occupations in the United States has shifted since ChatGPT’s release in November 2022. It compared that shift with earlier periods: the arrival of personal computers in the 1980s and the internet from the mid-1990s to the early 2000s.

The result was that the occupational mix is changing slightly faster than in those earlier periods, but “not a large difference.” The researchers found the shift was only about one percentage point above the internet-era path, and that some shifts began in 2021, before generative AI went mainstream. The share of workers in high-, medium- and low-exposure occupations stayed roughly stable.

Equally important was what the authors said about their own limits. The analysis is not predictive, cannot establish cause, and early effects might be too small to detect. The theoretical exposure scores used in it reflect what could be sped up, not what is. They also noted that major technological transitions have historically unfolded over decades, not a couple of years.

Are entry-level workers being hit harder?

A study from Stanford’s Digital Economy Lab, presented in a September 2025 seminar by Bharat Chandar and Ruyu Chen, argues yes. Using high-frequency payroll records from the largest payroll software provider in the United States, they reported a 13 percent relative decline in employment for workers aged 22 to 25 in the most AI-exposed occupations since generative AI spread widely, even after controlling for firm-level shocks.

The pattern was specific. Employment for less exposed fields, and for more experienced workers in the same exposed occupations, held steady or kept growing. The declines were concentrated where AI is more likely to automate work than to assist it, and the adjustment showed up in jobs rather than pay. The results held when technology firms and remote-friendly occupations were excluded.

The Yale team saw a milder hint of something similar, noting that the occupational mix of recent graduates has changed slightly faster lately. But they stressed that this could reflect AI, a weak labor market for new graduates, or both, and that their sample was small.

So the two studies are not exactly contradictory. One finds little change across the whole economy. The other finds a concentrated effect on a narrow group. A small share of workers can face real disruption without moving national averages.

Which kinds of tasks are most affected?

The evidence so far points to tasks that are digital, text-based and relatively routine in structure: drafting standard documents, summarizing, writing first-pass code, answering common customer questions, and searching for information. A 2025 preprint by Tomlinson and colleagues, built on how people used Microsoft Copilot, identified occupations whose tasks overlap heavily with what the tool does, including roles built around information gathering and communication.

That paper drew pushback too. One historian noted that the method reduced a historian’s work to retrieving and conveying information, which inflates the apparent overlap. It illustrates a general issue: a job is a bundle of tasks plus judgment, relationships, accountability and physical presence, and a method that scores only the easily describable tasks will overstate how replaceable the whole role is.

Work that depends on physical dexterity in unpredictable settings, such as skilled trades and hands-on care, has so far been less affected by language-focused AI. That does not guarantee it stays that way as robotics develops.

Has new technology destroyed jobs before?

Past waves of automation shifted work around more than they eliminated it. Computers removed typing pools and created whole industries. But the transitions were not painless: individual workers in declining roles often struggled to move, and regions took years to adjust. Economists such as Daron Acemoglu and Simon Johnson have argued that outcomes depend on how the technology is directed, whether toward replacing workers or toward helping them do more, and on policy and bargaining power.

The historical record is a reason for some caution about doom forecasts. It is not a law of nature. The current technology differs from earlier ones in being general-purpose and aimed at cognitive work, and nobody knows for certain whether the old pattern will repeat.

What is likely to change first?

Several things seem plausible, and they are labeled predictions, not findings. Entry-level hiring in fields with automatable first-draft work may tighten, which would matter because those jobs are how people build experience. Job descriptions may shift toward reviewing and directing AI output. Small teams may do work that once needed larger ones.

What can be said with more confidence is that the effect will not be uniform. Industry, task mix, company size, regulation and how quickly firms redesign workflows will all influence the outcome. Surveys of employers show many expect automation to cut some jobs, while data on platforms and big data suggest more neutral employment effects.

Why is it so hard to measure, and what else is at stake?

Part of the difficulty is data. The Yale analysis leaned on theoretical exposure scores from OpenAI, which describe what could be sped up rather than what is, and on usage data from Anthropic’s Claude, which covers only that product’s users and skews toward coding and writing. How the many tasks nobody observed are treated changes the automation share sharply. The authors called for fuller usage data from all the major AI companies.

Timing is another problem. Firms need months or years to rebuild workflows around a new tool, and a hiring freeze is invisible until you compare cohorts. Meanwhile other forces are moving at once: interest rates, post-2020 over-hiring, shifts in consumer spending. Pulling out an AI signal from that noise is the central challenge of the research, and it is why two careful studies can sound so different.

Employment counts are not the only issue. The OECD published a 2022 report on the main ethical risks of AI in the workplace, calling for governance frameworks and accountability around privacy, fairness and surveillance. Tools that score applicants, monitor productivity or assign shifts raise questions that have little to do with whether a job disappears. Earlier waves of robotics, for what it is worth, were associated in European studies with net employment growth overall, even as some businesses expected automation to cut specific roles.

What can individual workers do?

Nothing here is a guarantee, and none of it is personalized advice. Still, the research suggests some sensible directions. Learning to use the tools in your field is low-cost. Skills that combine judgment with context, such as client relationships, domain expertise and the ability to check and correct machine output, appear harder to replace than isolated tasks. Following your own industry’s hiring data, rather than general headlines, gives a better read of local conditions.

Common questions

Will AI replace most jobs?

No study has shown that. Exposure estimates say many jobs contain tasks AI can do, but actual employment data through 2025 show limited economy-wide change. Long-term outcomes are uncertain.

Which jobs are safest?

No job is guaranteed. Work that mixes physical presence, unpredictable environments and human interaction has been less affected so far. Roles built mainly around producing routine digital text have more overlap with current tools.

Is AI already causing layoffs?

Some companies have attributed cuts to AI, and large technology layoffs since 2023 have been widely reported. Separating AI from other causes, such as interest rates and over-hiring after 2020, is difficult, and researchers have not established how much to attribute to AI.

Should students avoid certain careers?

The evidence does not support a blanket rule. Entry-level hiring in some exposed fields looks softer, according to the Stanford paper, but the findings are early. Career decisions are better made with current local data and the guidance of advisers.

Where can I find reliable numbers?

Government labor statistics, central bank and IMF research, and university labor-economics centers publish methodology along with their results. Be cautious with consultancy reports that give a single dramatic figure without describing how it was built.

What to remember

Exposure is not the same as job loss. Through 2025, broad U.S. data show limited change in the overall occupational mix, while one large payroll study finds a concentrated decline among young workers in highly exposed roles. Both can be true. The long run is uncertain, history offers hope with caveats, and the most reliable approach is to follow evidence that comes with its methods attached.

Further reading

Economists' book-length argument on how technology choices shape who benefits. See it on Amazon (affiliate link)

As an Amazon Associate I earn from qualifying purchases. A commission may be earned if you buy through this link, at no extra cost to you.

Sources