A brand new measure and early proof Anthropic

Key Findings

• We introduce a brand new measure of AI displacement danger, noticed publicity, that mixes theoretical LLM functionality and real-world utilization knowledge, weighting automated (quite than augmentative) and work-related makes use of extra closely
• AI is way from reaching its theoretical functionality: precise protection stays a fraction of what is possible
• Occupations with greater noticed publicity are projected by the BLS to develop much less by 2034
• Workers in essentially the most uncovered professions usually tend to be older, feminine, extra educated, and higher-paid
• We discover no systematic improve in unemployment for extremely uncovered employees since late 2022, although we discover suggestive proof that hiring of youthful employees has slowed in uncovered occupations

Introduction

The speedy diffusion of AI is producing a wave of analysis measuring and forecasting its impacts on labor markets. But the monitor file of previous approaches offers purpose for humility.

For instance, a distinguished try to measure job offshorability recognized roughly 1 / 4 of US jobs as weak, however a decade on, most of these jobs maintained wholesome employment development. The authorities’s personal occupational development forecasts, whereas directionally right, have added little predictive worth past linear extrapolation of previous traits. Even in hindsight, the affect of main financial disruptions on the labor market is commonly unclear. Studies on the employment results of commercial robots attain opposing conclusions, and the dimensions of job losses attributed to the China commerce shock continues to be debated.¹

In this paper, we current a brand new framework for understanding AI’s labor market impacts, and check it towards early knowledge, discovering restricted proof that AI has affected employment to this point. Our objective is to ascertain an strategy for measuring how AI is affecting employment, and to revisit these analyses periodically. This strategy will not seize each channel by which AI may reshape the labor market, however by laying this groundwork now, earlier than significant results have emerged, we hope future findings will extra reliably establish financial disruption than post-hoc analyses.

It is feasible that the impacts of AI can be unmistakable. This framework is most helpful when the results are ambiguous—and will assist establish essentially the most weak jobs earlier than displacement is seen.

Counterfactuals

Causal inference is less complicated when the results are giant and sudden. The COVID-19 pandemic and accompanying coverage measures precipitated financial disruption so stark that subtle statistical approaches have been pointless for a lot of questions. For instance, unemployment jumped sharply within the early weeks of the pandemic, leaving little room for different explanations.

The impacts of AI, nonetheless, may be much less like COVID and extra just like the web or commerce with China. The results might not be instantly clear from mixture unemployment knowledge; components like commerce coverage and the enterprise cycle may cloud interpretations of development strains.

One frequent strategy is to check outcomes between roughly AI-exposed employees, companies, or industries, to be able to isolate the impact of AI from confounding forces.² Exposure is usually outlined on the activity degree: AI can grade homework however not handle a classroom, for instance, so academics are thought-about much less uncovered than employees whose total job will be carried out remotely.

Our work follows this task-based strategy, incorporating measures of theoretical AI functionality and real-world utilization, earlier than aggregating to occupations.³

Measuring publicity

Our strategy combines knowledge from three sources.

1. The O*NET database, which enumerates duties related to round 800 distinctive occupations within the US.
2. Our personal utilization knowledge (as measured within the Anthropic Economic Index).
3. Task-level publicity estimates from Eloundou et al. (2023), which measure whether or not it’s theoretically potential for an LLM to make a activity at the very least twice as quick.

Eloundou et al.’s metric, β, scores duties on a easy scale: 1 if a activity will be doubled in pace by an LLM alone, 0.5 if it requires extra instruments or software program constructed on prime of the LLM, and 0 in any other case.⁴

Why would possibly precise utilization fall wanting theoretical functionality? Some duties which are theoretically potential might not present up in utilization due to mannequin limitations. Others could also be sluggish to diffuse on account of authorized constraints, particular software program necessities, human verification steps, or different hurdles. For instance, Eloundou et al. mark “Authorize drug refills and provide prescription information to pharmacies” as absolutely uncovered (β=1). We haven’t noticed Claude performing this activity, though the evaluation appears right in that it may theoretically be sped up by an LLM.

That mentioned, these measures of theoretical functionality and precise utilization are extremely correlated. As Figure 1 reveals, 97% of the duties noticed throughout the earlier 4 Economic Index experiences fall into classes rated as theoretically possible by Eloundou et al. (β=0.5 or β=1.0).

A brand new measure of occupational publicity

Our new measure, noticed publicity, is supposed to quantify: of these duties that LLMs may theoretically pace up, which are literally seeing automated utilization in skilled settings? Theoretical functionality encompasses a wider vary of duties. By monitoring how that hole narrows, noticed publicity gives perception into financial modifications as they emerge.

Our measure qualitatively captures a number of facets of AI utilization that we expect are predictive of job impacts. A job’s publicity is greater if:

• Its duties are theoretically potential with AI
• Its duties see important utilization within the Anthropic Economic Index⁵
• Its duties are carried out in work-related contexts
• It has a comparatively greater share of automated use patterns or API implementation
• Its AI-impacted duties make up a bigger share of the general function⁶

We give mathematical particulars within the Appendix. We depend duties which are theoretically succesful with an LLM as lined if they’ve seen ample work-related utilization in Claude visitors. We then regulate for the way the duty is being carried out: absolutely automated implementations obtain full weight, whereas augmentative use receives half weight. Finally, the task-level protection measures are averaged to the occupation degree weighted by the fraction of time spent on every activity.

Figure 2 reveals noticed publicity (in crimson) in comparison with β from Eloundou et al. (in blue), illustrating the distinction between theoretical and precise use on our platform, grouped by broad occupational classes. We calculate this by first averaging to the occupation degree weighting by our time fraction measure, then averaging to the occupation class weighting by complete employment. For instance, the β measure reveals scope for LLM penetration within the majority of duties in Computer & Math (94%) and Office & Admin (90%) occupations.

The crimson space, depicting LLM use from the Anthropic Economic Index, reveals how individuals are utilizing Claude in skilled settings. The protection reveals AI is way from reaching its theoretical capabilities. For occasion, Claude at present covers simply 33% of all duties within the Computer & Math class.

As capabilities advance, adoption spreads, and deployment deepens, the crimson space will develop to cowl the blue. There is a big uncovered space too; many duties, in fact, stay past AI’s attain—from bodily agricultural work like pruning bushes and working farm equipment to authorized duties like representing shoppers in court docket.

Figure 3 reveals the ten occupations most uncovered beneath this measure. In line with different knowledge displaying that Claude is extensively used for coding, Computer Programmers are on the prime, with 75% protection, adopted by Customer Service Representatives, whose fundamental duties we more and more see in first-party API visitors. Finally, Data Entry Keyers, whose major activity of studying supply paperwork and getting into knowledge sees important automation, are 67% lined.

At the underside finish, 30% of employees have zero protection, as their duties appeared too occasionally in our knowledge to satisfy the minimal threshold. This group contains, for instance, Cooks, Motorcycle Mechanics, Lifeguards, Bartenders, Dishwashers, and Dressing Room Attendants.

How publicity tracks with projected job development and employee traits

The US Bureau of Labor Statistics (BLS) publishes common employment projections, with the most recent set, printed in 2025, overlaying predicted modifications in employment for each occupation from 2024 to 2034. In Figure 4, we evaluate our job-level protection measure to their predictions.

A regression on the occupation degree weighted by present employment finds that development projections are considerably weaker for jobs with extra noticed publicity. For each 10 proportion level improve in protection, the BLS’s development projection drops by 0.6 proportion factors. This gives some validation in that our measures monitor the independently derived estimates from labor market analysts, though the connection is slight. Interestingly, there isn’t any such correlation utilizing the Eloundou et al. measure alone.

Figure 5 reveals traits of employees within the prime quartile of publicity and the 30% of employees with zero publicity within the three months earlier than ChatGPT was launched, August to October 2022, utilizing knowledge from the Current Population Survey.⁷ The teams are very completely different. The extra uncovered group is 16 proportion factors extra more likely to be feminine, 11 proportion factors extra more likely to be white, and nearly twice as more likely to be Asian. They earn 47% extra, on common, and have greater ranges of training. For instance, folks with graduate levels are 4.5% of the unexposed group, however 17.4% of essentially the most uncovered group, an nearly fourfold distinction.

Prioritizing outcomes

With these publicity measures in hand, the query is what to search for. Researchers have taken completely different approaches. For instance, Gimbel et al. (2025) monitor modifications within the occupational combine utilizing the Current Population Survey. Their argument is that any essential restructuring of the financial system from AI would present up as modifications in distribution of jobs.¹ (They discover that, to this point, modifications have been unremarkable.) Brynjolfsson et al. (2025) have a look at employment ranges break up by age group utilizing knowledge from the payroll processing agency ADP, whereas Acemoglu et al. (2022) and Hampole et al. (2025) use job posting knowledge from Burning Glass (now Lightcast) and Revelio, respectively.

We deal with unemployment as our precedence consequence as a result of it most straight captures the potential for financial hurt—a employee who’s unemployed needs a job and has not but discovered one. In this case, job postings and employment don’t essentially sign the necessity for coverage responses; a decline in job postings for a extremely uncovered function could also be counteracted by elevated openings in a associated one. Most dangerous labor market developments of AI ought to arguably embody a interval of elevated unemployment, as displaced employees seek for options. The Current Population Survey is nicely suited to monitoring this, as unemployed respondents report their earlier job and business.

Initial outcomes

We subsequent research traits in unemployment, matching our occupation-level measures to respondents within the Current Population Survey.

A key query in decoding our protection measure is which employees ought to be thought-about handled? Should modifications in employment be anticipated from simply 10% activity protection? Gans and Goldfarb (2025) present that if an O-ring mannequin finest describes jobs, employment results may be seen solely when all duties have some extent of AI penetration. Hampole et al. (2025) argue that imply publicity decreases labor demand, however focus of publicity in solely sure duties can counteract this. And Autor and Thompson (2025) spotlight the extent of experience required for the remaining duties.

With a watch towards simplicity, and noting that we’re most involved with giant impacts, we middle our evaluation on the concept that impacts ought to be felt most within the teams with the very best imply publicity. We evaluate employees within the prime quartile of time-weighted activity protection to these within the backside. If AI capabilities advance shortly, activity protection may be excessive for decrease percentiles of protection, which could make an absolute threshold extra useful. But we make the belief that impacts ought to have an effect on essentially the most uncovered employees first, and current outcomes various the cutoff we use to outline remedy.

The higher panel of Figure 6 reveals uncooked traits within the unemployment price since 2016 for employees within the prime quartile of publicity and the unexposed group. During COVID, the much less AI-exposed employees—who usually tend to have in-person jobs—noticed a a lot bigger improve in unemployment. Since then, the traits have been largely related between the 2 teams. The decrease panel measures the scale of the hole between essentially the most and least uncovered employees in a difference-in-differences framework, mirroring the findings from the uncooked knowledge. The common change within the hole because the launch of ChatGPT is small and insignificant, suggesting that the unemployment price of the extra uncovered group has elevated barely however the impact is indistinguishable from zero.⁸

What form of situations can this framework establish? Based on the boldness interval of the pooled estimate, differential will increase in unemployment on the order of 1 proportion level can be detectable (it will change as new knowledge is available in, so it’s merely a ballpark estimate). If all employees throughout the prime 10% have been laid off, it might improve unemployment throughout the prime quartile group from 3% to 43%, and it might improve mixture unemployment from 4% to 13%.

A smaller however nonetheless regarding affect can be a state of affairs corresponding to a “Great Recession for white-collar workers.” During the 2007-2009 Great Recession, unemployment charges doubled from 5% to 10% within the US. Such a doubling within the prime quartile of publicity would improve its unemployment price from 3% to six%. This ought to be seen in our evaluation as nicely. Note that our core estimate relies on differential modifications within the unemployment price within the uncovered group in comparison with the much less uncovered group. If unemployment elevated for all employees in parallel, we’d not attribute this to AI developments that also depart many duties unaffected.

One group of explicit concern is younger employees. Brynjolfsson et al. report a 6—16% fall in employment in uncovered occupations amongst employees aged 22 to 25. They attribute this lower primarily to a slowdown in hiring quite than a rise in separations.⁹

We discover that the unemployment price for younger employees within the uncovered occupations is flat (see Appendix). But slowed hiring might not essentially manifest as elevated unemployment, since many younger employees are labor market entrants and not using a listed occupation within the CPS knowledge and should exit the labor drive quite than seem as unemployed. To tackle hiring straight, we use the panel dimension of the CPS, counting the p.c of younger (22-25 yr outdated) employees who start a brand new job in a extra vs. much less uncovered occupation over time. Figure 7 reveals the month-to-month job discovering price (i.e., when a employee experiences a job that they didn’t have within the earlier month) for younger employees, break up by whether or not they’re getting into a high- vs. low-exposure occupation.

Apart from some giant swings in 2020-2021, these sequence visually diverge in 2024, with younger employees comparatively much less more likely to be employed into uncovered occupations. Job discovering charges on the much less uncovered occupations stay steady at 2% per thirty days, whereas entry into essentially the most uncovered jobs decreases by about half a proportion level. The averaged estimate within the post-ChatGPT period is a 14% drop within the job discovering price in comparison with that in 2022 within the uncovered occupations, though that is simply barely statistically important. (There is not any such lower for employees older than 25.)

This might present some sign of the early results of AI on employment, and echoes the findings from Brynjolfsson et al. But there are a number of different interpretations. The younger employees who should not employed could also be remaining at their present jobs, taking completely different jobs, or returning to high school. An additional data-related caveat is that job transitions could also be extra weak to mismeasurement in surveys.¹⁰

Discussion

This report introduces a brand new measure for understanding the labor market results of AI and research impacts on unemployment and hiring. Jobs are extra uncovered to AI to the extent that their duties are theoretically possible with LLMs and noticed on our platforms in automated, work-related use circumstances. We discover that pc programmers, customer support representatives, and monetary analysts are among the many most uncovered. Using survey knowledge from the US, we discover no affect on unemployment charges for employees in essentially the most uncovered occupations, though there’s tentative proof that hiring into these professions has slowed barely for employees aged 22-25.

Our work is a primary step towards cataloging the affect of AI on the labor market. We hope that the analytical steps taken on this report, particularly round protection and counterfactuals, can be simple to replace as new knowledge on employment and AI utilization emerge. An established strategy might assist future observers separate sign from noise.

There are a number of enhancements to be made to the current work. Our utilization knowledge can be included in future updates, forming an evolving image of activity and job protection within the financial system. The Eloundou et al. metric may be up to date, to the extent that it’s linked to LLM capabilities as of early 2023. And, given the suggestive outcomes round younger employees and labor market entrants, a key subsequent step may be to take a look at how latest graduates with academic credentials in uncovered areas are navigating the labor market.

Appendix

Available here.

Acknowledgements

Written by Maxim Massenkoff and Peter McCrory.

With acknowledgements to: Ruth Appel, Tim Belonax, Keir Bradwell, Andy Braden, Dexter Callender III, Miriam Chaum, Madison Clark, Jake Eaton, Deep Ganguli, Kunal Handa, Ryan Heller, Lara Karadogan, Jennifer Martinez, Jared Mueller, Sarah Pollack, David Saunders, Carl De Torres, Kim Withee, and Jack Clark.

We moreover thank Martha Gimbel, Anders Humlum, Evan Rose, and Nathan Wilmers for suggestions on earlier variations of this report.

Citation

@on-line{massenkoffmccrory2026labor,
 writer = {Maxim Massenkoff and Peter McCrory},
 title = {Labor market impacts of AI: A brand new measure and early proof},
 date = {2026-03-05},
 yr = {2026},
 url = {https://www.anthropic.com/research/labor-market-impacts},
}

References

Acemoglu, Daron and Pascual Restrepo, “Robots and Jobs: Evidence from US Labor Markets,” Journal of Political Economy, 2020, 128 (6), 2188–2244.

Acemoglu, Daron, David Autor, Jonathon Hazell, and Pascual Restrepo, “Artificial intelligence and jobs: Evidence from online vacancies,” Journal of Labor Economics, 2022, 40 (S1), S293–S340.

Appel, Ruth, Maxim Massenkoff, Peter McCrory, Miles McCain, Ryan Heller, Tyler Neylon, and Alex Tamkin, “Anthropic Economic Index report: economic primitives,” 2026.

Autor, David H, David Dorn, and Gordon H Hanson, “The China syndrome: Local labor market effects of import competition in the United States,” American Economic Review, 2013, 103 (6), 2121–2168.

Autor, David H, & Thompson, N. (2025). Expertise. NBER Working Paper, (w33941).

Blinder, Alan S et al., “How many US jobs might be offshorable?,” World Economics, 2009, 10 (2), 41.

Borusyak, Kirill, Peter Hull, and Xavier Jaravel, “Quasi-experimental shift-share research designs,” The Review of Economic Studies, 2022, 89 (1), 181–213.

Brynjolfsson, Erik, Bharat Chandar, and Ruyu Chen, “Canaries in the coal mine? six facts about the recent employment effects of artificial intelligence,” Digital Economy, 2025.

Eckhardt, Sarah and Nathan Goldschlag, “AI and Jobs: The Final Word (Until the Next One),” Economic Innovation Group (EIG), August 2025. Available at: https://eig.org/ai-and-jobs-the-final-word/

Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock, “Gpts are gpts: An early look at the labor market impact potential of large language models,” arXiv preprint arXiv:2303.10130, 2023, 10.

Fujita, S., Moscarini, G., & Postel-Vinay, F. (2024). Measuring employer-to-employer reallocation. American Economic Journal: Macroeconomics, 16(3), 1-51.

Gans, Joshua S. and Goldfarb, Avi, “O-Ring Automation,” NBER Working Paper No. 34639, December 2025. Available at SSRN: https://ssrn.com/abstract=5962594

Gimbel, Martha, Molly Kinder, Joshua Kendall, and Maddie Lee, “Evaluating the Impact of AI on the Labor Market: Current State of Affairs,” Research Report, The Budget Lab at Yale, New Haven, CT October 2025. Available at: https://budgetlab.yale.edu.

Graetz, Georg and Guy Michaels, “Robots at Work,” Review of Economics and Statistics, 2018, 100 (5), 753–768.

Hampole, Menaka, Dimitris Papanikolaou, Lawrence DW Schmidt, and Bryan Seegmiller, “Artificial intelligence and the labor market,” Technical Report, National Bureau of Economic Research 2025.

Handa, Kunal, Alex Tamkin, Miles McCain, Saffron Huang, Esin Durmus, Sarah Heck, Jared Mueller, Jerry Hong, Stuart Ritchie, Tim Belonax, Kevin Ok. Troy, Dario Amodei, Jared Kaplan, Jack Clark, and Deep Ganguli, “Which Economic Tasks are Performed with AI? Evidence from Millions of Claude Conversations,” 2025.

Hui, Xiang, Oren Reshef, and Luofeng Zhou, “The short-term effects of generative artificial intelligence on employment: Evidence from an online labor market,” Organization Science, 2024, 35 (6), 1977–1989.

Johnston, Andrew and Christos Makridis, “The labor market effects of generative AI: A difference-in-differences analysis of AI exposure,” Available at SSRN 5375017, 2025.

Massenkoff, Maxim, “How predictable is job destruction? Evidence from the Occupational Outlook,” 2025. Working Paper.

Ozimek, Adam, “Overboard on Offshore Fears,” 2019. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3777307

Tamkin, Alex and Peter McCrory, “Estimating AI productivity gains from Claude conversations,” 2025.

Tomlinson, Ok., Jaffe, S., Wang, W., Counts, S., & Suri, S. (2025). Working with AI: measuring the applicability of generative AI to occupations. arXiv preprint arXiv:2507.07935.

Footnotes

1. Job offshorability: Blinder et al. (2009) and Ozimek (2019); Government development forecasts: Massenkoff (2025); Robots: Graetz and Michaels (2018) and Acemoglu and Restrepo (2020); China shock: Autor et al. (2013) and Borusyak et al. (2022).

2. Brynjolfsson et al. (2025) evaluate employment traits for employees in additional versus much less AI-exposed occupations, utilizing the duty publicity measures from Eloundou et al. (2023) and payroll knowledge from ADP. Johnston and Makridis (2025) do an analogous task-based evaluation utilizing US administrative knowledge, however they mixture remedy to the business degree. Hui et al. (2024) research how freelance jobs on Upwork responded to the discharge of ChatGPT and superior picture technology instruments, evaluating employees in straight affected classes to these in unaffected classes earlier than and after every instrument’s launch date. Hampole et al. (2025) instrument for firm-level AI adoption utilizing historic college hiring networks: companies that traditionally recruited from universities whose graduates later entered AI-related roles confronted decrease adoption prices.

3. Our task- and occupation-level publicity measures can readily incorporate different utilization knowledge, and be prolonged to completely different nations. We intend to use this technique to new settings over time.

4. In their framework, “Directly exposed’” duties have been people who could possibly be accomplished in half the time with an LLM (with a 2,000-word enter restrict and no entry to latest information). Tasks that have been “exposed with tools” have been these topic to the identical speedup with an LLM that had entry to software program for, e.g., info retrieval and picture processing. Tasks that weren’t uncovered couldn’t have their length lowered by 50% or extra utilizing an LLM.

5. We use the earlier two Anthropic Economic Index datasets, overlaying utilization from August and November 2025. For ONET duties which are extremely semantically related, we break up the counts throughout them.

6. There are judgment calls concerned at each step. Should the Eloundou et al. (2023) measure enter as {0, 0.5, 1} or one thing else? What determines “significant” use? How can we deal with duties which appear similar to these with excessive utilization, however are too uncommon to have been picked up particularly within the sampling for the Economic Index? How far more ought to automation workflows depend in comparison with augmentation? A reassuring discovering which we develop on within the Appendix is that the Spearman (rank-rank) correlation of job publicity throughout many resolutions to those questions is exceedingly excessive.

7. To match O*NET-SOC codes to occ1990 codes within the CPS, we use the crosswalk supplied by Eckhart and Goldschlag (2025).

8. We discover this additional in 3 ways within the Appendix. First, we ask whether or not the percentile cutoff that we use to outline remedy issues, various it from the median to the ninety fifth percentile. In all circumstances, the affect is flat or unfavourable (which means that unemployment decreases for the uncovered group). Next, we deal with younger employees specifically, these aged 22 to 25 as in Brynjolfsson et al. (2025). Finally, we use knowledge on unemployment insurance coverage claimants from the Department of Labor to measure the unemployment, quite than CPS survey responses. In no extension do we discover clear impacts on uncovered jobs.

9. This vary is extensive as a result of the authors present estimates towards a number of counterfactuals. The 6 proportion level drop compares to a counterfactual of flat employment development. The 16 proportion level estimate comes from a design evaluating related employees in the identical agency with completely different occupations.

10. See Fujita, et al. (2024).