A map for AI’s labor market impacts

01

A recurring fear and ambiguous outcomes

01

First, a quick historical refresher: the fear that a new technology would displace swaths of labor is not new.

This is the reason that led the Luddites to smash looms in the 1810s (Frey 2019). Around a century later, telephone switchboard operating, at that time among the largest occupations for young women in America, all but disappeared when automatic dial switching arrived. In the 1970s, the introduction of ATMs could have meant the same fate for bank tellers. And most recently, in 2016, the “godfather of AI” Geoffrey Hinton suggested that hospitals should “stop training radiologists now,” since deep learning would outperform them within five years (Creative Destruction Lab 2016).

What happened then? Despite the disappearance of the telephone operator occupation, the overall employment of subsequent cohorts of women did not plummet as they moved into other sectors including completely new categories of work (Feigenbaum and Gross 2024).

More surprisingly, the number of bank tellers has actually doubled since the introduction of the ATM. One of the reasons was that it became cheaper for banks to run branches and consequently expanded, opening more branches, ultimately requiring more bank tellers (Bessen 2015). And radiology has so far followed the pattern of the bank tellers, not the telephone operators. One reason is that radiologists do not just read scans but do so much more, including weighing the image against the patient’s history, communicating with colleagues and patients, drawing conclusions, and making decisions based on all of that. Thus, while software might take over part of that work, the rest does not vanish but moves to the foreground (Agrawal, Gans and Goldfarb 2019).

What can we learn from these historical phenomena for the labor market effects of AI? In particular, can they help explain why some scholars expect labor displacement to dominate, while others stress that workers will be better off or at least not worse off on balance?

02

An economic map

02

We first need a simple way to describe what technology does to a job.

Economists would call such a description a model. A model is always a simplified picture of reality, often an offensively simple one, so it is fair to ask why we would describe the world with something that knowingly leaves almost all of it out. The reason is that a model is built to answer a particular question.

Think of a map: when you are hiking, a good one shows where to turn and which viewpoints are worth the climb, and it earns that usefulness precisely by leaving out every single tree, which would only crowd the page and bury the route. An economic model works the same way: for a given question, it keeps the few elements that decide the answer and the mechanisms worth seeing, and sets the rest aside. (Rodrik 2015)

To think about AI’s effects on jobs, economists most often reach for what they call the task-based model (Acemoglu, Autor 2011). Its starting point is that any job can be broken down into many distinct tasks. As with the radiologist, each thing a worker does is a task, and the job is the bundle of them. The framework then asks, task by task, whether it is done by a person or by a machine. As technology improves, it becomes possible to automate tasks once done by people.

While this means that one task might shrink or disappear, the tasks left in human hands can matter more, provided they complement the one now automated. This is called augmentation, and it is the radiologists’ case: once machines handle the reading, the judgment and the decision built on top of it become the scarce and thus valuable part of the job.

At the same time, new technology gives rise to completely novel tasks and jobs as seen when many women moved into industries other than telephone operation. On top of that, the rise in overall productivity might lift wages, raise revenues, lower costs, or a mix of all of them. In the case of ATMs and bank branches, the lower costs played a crucial role, leading to a net increase in the number of bank tellers.

03

Sources of the disagreement

Empirical vs. Methodological

03 – EMPIRICAL

The empirical is about the distance between the framework and the data. To start, economists rely on databases such as O*NET which divides the entire workforce into nearly 1,000 occupations and close to 20,000 task descriptions. Crucially, though, whether a task is automatable is not the same as it being automated. One line of research measures technical feasibility, the tasks that today’s models could in principle perform (Eloundou et al. 2024); another measures actual use, from records of how people interact with AI tools (Tomlinson et al. 2025). The two need not agree, and there is ongoing debate about this gap, as well as about where and how fast AI’s effects will show up.

For one, a task can be automatable and still be done by a person because the person does it more cheaply and thus it is not profitable for the firm to automate the task (Lindenlaub et al. 2026). For another, adopting AI tools does not necessarily change workflows, as this would require wider institutional changes such as new forms of team-level coordination (Dillon et al. 2026). Yet others stress that the effects arrive unevenly, landing differently on early-career and experienced workers (Hosseini & Lichtinger 2026; Teeselink 2025), on lower- and higher-skill and -experience jobs (Brynjolfsson, Li, Raymond 2025), on tasks implying some jobs’ expertise requirement falling while others’ expertise is becoming more valuable (Autor, Thompson 2025). The same technology can look very different in the data depending on where you look.

03 – METHODOLOGICAL

The methodological bucket highlights the role of certain conditions within the canonical framework or proposes new and currently overlooked mechanisms. One is the role of how different tasks within a job relate to each other and whether one could be substituted for another. The general view is that reading scans and interacting with patients are not substitutes for each other but rather complements. That means both are needed together for the radiologist to do a successful job.

On an economy-wide level, this pattern, however, implies that certain tasks, jobs and potentially even sectors that require human input could serve as bottlenecks. Even when a task, or a whole sector, is automated away, the money saved is spent somewhere, and labor tends to flow toward whatever machines have not made cheap, so the aggregate effect can stay modest even as particular jobs are upended (Jones, Tonetti 2026).

As outlined before, the task-based model prescribes that the overall effect is the net-outcome between new tasks and productivity gains on the one hand, and displacement on the other hand. If technology takes the form of small productivity effects with concomitant substantial displacement, the negative effects are likely to dominate. If these affected workers, however, can easily move to other tasks and jobs, the effect on affected workers again is less clear and there are further ripple effects on other indirectly affected workers (Acemoglu, Kong, Restrepo 2025).

Where does that leave us?

These are some reasons for the diverging predictions about AI’s impacts. On top of that, there are more speculative arguments emphasizing AI’s potential to displace more jobs than any previous technology, its unprecedented speed, distributional concerns and uncertainties surrounding AI reaching capabilities that are no longer understandable to humans.

Taking it all together, while this post cannot offer a definitive answer about AI’s future impact on jobs*, with new methodologies discussed and empirical findings uncovered, the debate remains lively and goes on.

References

Acemoglu, Daron, and David Autor. 2011. "Skills, tasks and technologies: Implications for employment and earnings." Handbook of labor economics. Vol. 4. Elsevier. 1043-1171.
Acemoglu, Daron, Linan Kong, and Pascual Restrepo. 2025. "Tasks at Work: Comparative Advantage, Technology, and Labor Demand." Handbook of Labor Economics 6: 1-114.
Agrawal, Ajay, Joshua S. Gans, and Avi Goldfarb. 2019. "Artificial Intelligence: The Ambiguous Labor Market Impact of Automating Prediction." Journal of Economic Perspectives 33 (2): 31–50.
Althoff, Lukas, and Hugo Reichardt. 2026. "Task-Specific Technical Change and Comparative Advantage." NBER Working Paper No. 35353.
Autor, David, and Neil Thompson. 2025. "Expertise." Journal of the European Economic Association 23 (4): 1203–1271.
Bessen, James. 2015. "Toil and Technology." Finance & Development 52 (1): 16–19. Washington, DC: International Monetary Fund.
Brynjolfsson, Erik, Danielle Li, and Lindsey R. Raymond. 2025. "Generative AI at Work." Quarterly Journal of Economics 140 (2): 889–942.
Creative Destruction Lab. 2016. "Geoff Hinton: On Radiology." YouTube video, https://www.youtube.com/watch?v=2HMPRXstSvQ (accessed 7/29/2026).
Dillon, Eleanor W., Sonia Jaffe, Nicole Immorlica, Christopher T. Stanton. Forthcoming. "Shifting Work Patterns with Generative AI." American Economic Review: Insights.
Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock. 2024. "GPTs are GPTs: Labor Market Impact Potential of LLMs." Science 384 (6702): 1306–1308.
Feigenbaum, James, and Daniel P. Gross. 2024. "Answering the Call of Automation: How the Labor Market Adjusted to Mechanizing Telephone Operation." Quarterly Journal of Economics 139 (3): 1879–1939.
Frey, Carl B. 2019. "The Technology Trap: Capital, Labor, and Power in the Age of Automation." Princeton: Princeton University Press.
Hosseini Maasoum, Seyed Mahdi, and Guy Lichtinger. 2025. "Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data". SSRN Working Paper No. 5425555, Harvard University.
Jones, Charles I., and Christopher Tonetti. 2026. "Past Automation and Future A.I.: How Weak Links Tame the Growth Explosion." Working paper (NBER).
Klein Teeselink, Bouke. 2025. "Generative AI and Labor Market Outcomes: Evidence from the United Kingdom." SSRN Working Paper No. 5516798, King's College London.
Lindenlaub, Ilse, Ryungha Oh, Maria Alejandra Rodriguez, and Laura Veldkamp. 2026. "Beyond Exposure: Predicting AI Adoption Based on Comparative Advantage." NBER Working Paper No. 35271.
O*NET. U.S. Department of Labor, Employment and Training Administration. https://www.onetonline.org/
Rodrik, Dani. 2015. Economics Rules: The Rights and Wrongs of the Dismal Science. New York: W. W. Norton & Company.
Tomlinson, Kiran, Sonia Jaffe, Will Wang, Scott Counts, and Siddharth Suri. 2025. "Working with AI: measuring the applicability of generative AI to occupations." arXiv preprint arXiv:2507.07935.
Trammell, Philip, and Anton Korinek. 2026. "Economic Growth under Transformative AI." NBER Working Paper No. 31815.

Author

A man with short hair, a beard, and glasses is smiling. He is wearing a suit, dress shirt, and tie. The image has a sepia-tone, sketch-like effect with a plain background.

Lukas Wiedemann is an economist in the Futures Team at Microsoft AI. His research interests include technological change and its implications for the structural transformation of the economy and society. Before his work at Microsoft AI, he conducted PhD research in Economics at the London School of Economics and Political Science.

EXPLORE

LINKS

LATEST

The architect as humanist builder: why AI can’t replace craft