AI and the Future of Work: Are Entry-Level Jobs Disappearing?
New evidence on how AI is actually used at work finds no unemployment shock among the most exposed workers, and none yet among the class of 2026. It finds something quieter instead: fewer young people getting in. What that means for emerging economies like India, for policy, and for anyone whose sense of self is bound up with a job.
No mass layoffs among exposed workers. But young people are being hired into exposed entry jobs 14% less often.
Two landmark 2026 studies, a new measure of observed AI use at work (Massenkoff & McCrory) and the first test of the graduating class of 2026 (Fairlie & Wu, UCLA/NBER), find no systematic rise in unemployment among AI-exposed workers, with the summer 2026 graduate unemployment rate holding steady at 7.3%. However, beneath the calm headline rate, the entry door is quietly narrowing: 22-to-25-year-olds are entering AI-exposed occupations 14% less often than in 2022, and the share of graduates who have stopped searching reached a five-year peak of 10.4%. For emerging economies like India where youth comprise 83% of the unemployed, the risk is not immediate job destruction, but a generation struggling to find the first rung on the career ladder.
AI and the Future of Work: Are Entry-Level Jobs Disappearing?
At a Glance: The Numbers That Matter
Unemployment Gap Shift
Difference between most & least exposed US workers since late 2022; statistically zero.
Young Job Entry Drop
Fall in rate at which 22–25 year olds start jobs in most exposed roles.
Summer 2026 Grad Jobless
US graduates aged 22–25, well within 6.3%–7.8% historical band.
Counting Sidelined Grads
Graduates wanting work who stopped searching; highest of 5 summers.
Actual vs Potential Use
Share of computer & maths tasks where AI is used vs what it could do.
Indian Tech Workforce
Tech services workforce by 2031 under BAU, vs +4M AI roles with skilling.
Prologue: The Question Beneath the Payslip
If a machine could do the work that names you, what would be left of the name?
Ask a stranger who they are and, within two sentences, most will tell you what they do. We are, it seems, what we are paid for. The Austrian social psychologist Marie Jahoda understood this better than most. In the early 1930s she studied Marienthal, a village near Vienna whose textile mill had shut, and watched men with nowhere to be lose their grip on the hours. Decades later she named what a job quietly supplies beyond a wage: a shape to the day, company outside the family, a share in some larger purpose, a place in the social order, and the plain discipline of regular activity [6]. Take the job away and all five go at once.
That is why the argument about artificial intelligence and work never stays economic for long. Sooner or later it arrives at a particular desk. A 23 year old in Pune, one of the fastest growing technology hubs in India, rewriting her CV for the fortieth analyst opening. A support executive in Gurugram who has noticed that the easy tickets no longer reach him. An accountant in Kolkata, twelve years into her career, wondering whether the skill she built is about to become cheap.
The public mood swings between alarm and dismissal. Bill Gates has written that he is “especially worried about young people, who will enter a workforce with fewer entry-level openings”; Jensen Huang, the CEO of chip giant NVIDIA, calls talk of AI reducing jobs “complete nonsense” [7], [8]. Where evidence is thin, anxiety rushes in. So the more useful question is not whether AI will take our jobs, but something narrower and more honest: what, today, can we actually measure about who is being touched, and how?
Two landmark empirical studies published in 2026 offer the most careful answers yet. In March 2026, economists Maxim Massenkoff and Peter McCrory at the AI research lab Anthropic (the creators of Claude) published a new study analyzing which jobs AI is actually being used in [1], drawing directly from the newly released Anthropic Economic Index. In September 2026, Robert Fairlie and Jane Wu of UCLA tested whether the graduating class of 2026—the very first cohort to finish their degrees in a workplace where generative AI tools were already ubiquitous—struggled to find work [2].
Their headlines are calmer than the apocalyptic predictions: no systematic surge in unemployment. Their fine print is sharper. The door into exposed jobs appears to be narrowing, and it is narrowing specifically for the young.
1. A Century of False Alarms, and a Better Instrument
In the autumn of 1930, with the world sliding into depression, John Maynard Keynes warned of a “new disease”: technological unemployment, which he described as “our discovery of means of economising the use of labour outrunning the pace at which we can find new uses for labour” [9]. The diagnosis has been repeated in every decade since. It has rarely been confirmed.
Forecasters have earned their humility. A widely cited study once judged roughly a quarter of US jobs to be offshorable; ten years on, most of those occupations were still adding workers [10], [11]. Official US occupational projections, though pointed in the right direction, have added little beyond extending past trends [12]. Even in hindsight, economists still argue over how many jobs industrial robots or Chinese import competition actually destroyed.
From “Could” to “Does”: Introducing the Anthropic Economic Index
Part of the trouble lies in what researchers measure. Most early AI exposure scores asked what a model theoretically could do on paper, based on computer science evaluations of occupational task lists from O*NET (the standard US Department of Labor occupational database).
Massenkoff and McCrory [1] took a different approach. In March 2026, Anthropic released the Anthropic Economic Index, an empirical dataset analyzing millions of anonymized, work-related user interactions to see which specific workplace tasks professionals are actually delegating to large language models. They termed the result observed exposure².
How is a Job Scored for AI Exposure?
Observed exposure measures what AI is actually used to do on the job (tracked via the Anthropic Economic Index from millions of real enterprise interactions), rather than speculative theoretical capability.
Assisting = 0.5 pt
Task Distribution & Exposure Calculator
Under their method, a task counts as exposed only if a language model could at least double a worker’s speed [13] and if evidence of real workers using AI for that exact task turns up repeatedly in the Anthropic Economic Index. Fully automated use counts in full (1.0); augmentative use, where a person remains in the loop to review and sign off, counts half (0.5). Task scores are then averaged into each occupation according to how much working time each task takes in an average week.
Possibility and practice turn out to be close cousins. Tasks rated fully feasible account for 68% of observed usage, partly feasible tasks for 29%, and infeasible ones for only 3% [1].
2. The Distance Between Can and Does
The most revealing number in the study is a gap. On paper, language models could speed up 94% of tasks in computer and mathematical occupations and 90% in office and administrative ones. In practice, observed use covers just 33% of computer and mathematical tasks [1]. The remainder is capability in waiting, held back by regulation, missing software, verification rules and, not least, trust.
Observed AI Workplace Exposure by Occupation (%)
At the other end, almost 30% of US workers register no observed coverage at all. They are mainly cooks, motorcycle mechanics, lifeguards, bartenders and dishwashers, people whose work is done with hands, in rooms, in real time.
The Gap is Starting to Close
Fairlie and Wu [2] argue that what matters now is not a new leap in raw model capability, but the gradual closing of this adoption gap, and 2026 is when it began to narrow in earnest. A US Census Bureau survey of more than 117,000 firms, fielded from November 2025 to February 2026, found that among firms using AI in place of employee tasks, the share replacing a large number of tasks rose from 2.5% to 7.0% since early 2024 [3].
Similarly, in the Ramp AI Index (a corporate spending benchmark tracking about 70,000 US companies) [14], median monthly AI spending per employee more than doubled between December 2025 and July 2026. Crucially, metered API use—the kind of automated software integration that embeds AI directly into enterprise backend systems, rather than employees individually typing queries into a web chatbot—surpassed half of all business AI spending.
A Faint but Telling Signal
The US Bureau of Labor Statistics projects employment for every occupation from 2024 to 2034 [15]. Occupations with higher observed exposure are expected to grow a little more slowly: each 10 percentage point rise in coverage goes with a projected growth rate 0.6 points lower. The link is weak, yet revealing, because the purely theoretical score shows no such relationship. Use, it seems, is a better early warning than possibility.
3. The New Face of Exposure
Earlier waves of automation came for the factory floor and the typing pool, the routine jobs in the middle of the wage ladder. This one is climbing higher. Using microdata from the Current Population Survey (CPS)³—the monthly household survey of more than 130,000 individuals conducted jointly by the US Census Bureau and Bureau of Labor Statistics—for August to October 2022 (the three months before ChatGPT launched), Massenkoff and McCrory [1] compare the most exposed quarter of workers with the 30% who have no exposure at all.
Demographic Profile: No Exposure vs. High Exposure
Comparing the composition of workers in the bottom quartile (no exposure) vs top quartile (high exposure) of AI impact
The contrast is stark. Exposed workers earn about 47% more an hour and are almost four times as likely to hold a graduate degree. They are also less than half as likely to belong to a union (5.3% against 11.7%), which leaves fewer collective buffers should disruption arrive. The people most exposed to AI are, in short, the people who were told that education was their insurance.
AI exposure now sits with the educated, the better paid and women, the groups least protected by unions.
4. No Layoff Wave, but a Narrowing Door
Unemployment: The Dog That Has Not Barked
If AI were pushing exposed workers out of their jobs, their unemployment rate should be rising faster than everyone else. It is not. Since late 2022 the gap between the most exposed and the unexposed has moved by an estimated 0.2 percentage points¹, too small to tell apart from zero. Nor is this a case of looking and failing to see. The econometric method would catch a differential rise of about 1 point, and a “Great Recession for white collar workers”, in which exposed unemployment doubled from roughly 3% to 6%, would stand out plainly [1].
Other researchers see the same calm. The Budget Lab at Yale finds that the occupational makeup of the US economy has shifted unremarkably since ChatGPT arrived [16].
Hiring: The Canaries in the Coal Mine
The disturbance, if there is one, is at the entrance rather than the exit. The Stanford economists Erik Brynjolfsson, Bharat Chandar and Ruyu Chen chose their title with care: Canaries in the Coal Mine? Using high-frequency payroll data, they found that employment among 22 to 25 year olds in exposed occupations fell by 6 to 16%, driven by fewer hires rather than more firings [17].
Among young workers aged 22 to 25, entry into the most AI-exposed occupations slipped by about half a percentage point, a 14% fall against 2022 levels, while entry into physical and low-exposure occupations held perfectly steady.
Massenkoff and McCrory hear an echo in survey data. Among young workers, the monthly rate of starting a job in a low exposure occupation holds steady at about 2%. Entry into the most exposed occupations has slipped by about half a percentage point, a 14% fall against 2022 levels. The estimate only just clears the bar of statistical significance⁴, and workers over 25 show no such decline.
The caveats deserve their due. Young people not hired into exposed roles may be staying put, taking other work or going back to study. Still, the direction is hard to ignore. AI seems to be sawing at the lowest rung of the ladder before it troubles anyone already standing on it.
AI is altering recruitment velocity before it alters termination rates.
Firms are not dismissing mid-career professionals whose institutional context and domain judgement remain vital. Instead, AI-augmented seniors are absorbing entry-level tasks, quietly reducing the number of junior requisitions opened each quarter.
5. The Class of 2026: The First Real Test
If AI is closing the entry door, summer 2026 is when it should show. Graduates enter the job market in a narrow window between June and August, and this was the first cohort to graduate after firms began using AI intensively. Robert Fairlie and Jane Wu followed US graduates aged 22 to 25 who hold a bachelor degree and are no longer studying, using Current Population Survey microdata from January 2022 to August 2026 [2].
Two choices make their test sharper than most. First, summer matters: graduate unemployment in June to August typically runs nearly 2 points above its level of about 5% in the rest of the year, so 2026 has to be compared with earlier summers. Second, they count the sidelined⁵: graduates who state they want a job and are ready to work, but have stopped submitting applications in the past four weeks (often due to discouragement), and so fall outside official headline unemployment rates.
Summer Unemployment of Young US Graduates (2022–2026)
US graduates aged 22 to 25 with bachelor degrees, June to August average. Comparing the official rate against graduates who want work but stopped searching (the sidelined).
The result is steadiness. Official summer unemployment was 7.3% in 2026, against 7.1% in 2022, 6.3% in 2023, 7.8% in 2024 and 7.2% in 2025. Adding the sidelined lifts 2026 to 10.4%, the highest of the five summers, but only 0.3 points above 2024. Regression estimates for summer 2026 are small and statistically insignificant in every specification, and fewer than 17% of unemployed recent graduates had lost a job through layoffs; most simply had not yet found one.
The Hints Beneath the Calm
How Young Graduates Fared
Comparing 2025 to 2026 labor market outcomes for recent graduates
None of these shifts is statistically significant, and part of the narrowing came from non graduates doing better rather than graduates doing worse. Other margins lean the same way. The share of young graduates either unemployed or working in jobs that need no formal qualification rose to 17.1% in summer 2026, from 15.6% in 2022.
AI, or the Empty Office?
When Fairlie and Wu split graduates by occupation, the link with AI exposure was positive but not significant, on both the observed and the theoretical measure. The clearer link was with remote work: graduates from occupations that can be done from home⁶ fared measurably worse.
The two overlap heavily, with correlations of 0.58 and 0.66 between remote work and the two AI measures, and earlier research finds remote work thins out the informal desk-side mentoring that junior staff need [18]. For now, AI and the empty office are hard to tell apart.
The class of 2026 did not hit a wall. But the measures that look past the official rate point the same way as the canaries: the entrance is a little narrower, and more of the young who cannot get in are waiting on the sidelines rather than being counted as unemployed.
6. Beyond America: The Global and Indian Picture
Wherever researchers look, the pattern rhymes: exposure is wide, displacement so far is thin, and the pressure gathers around clerical and early career work.
Global AI Exposure Estimates
International perspectives on AI's labor market impact
The International Labour Organization (ILO) is careful on this point. Because almost every occupation still contains tasks that need a human, the likeliest outcome is that most jobs will be transformed rather than erased [4].
What This Means for India
India stands at an awkward crossroads. Over three decades, its services export model has been anchored in precisely the tasks at the top of the exposure table: writing code, running automated quality assurance, answering customer queries and processing financial data.
NITI Aayog (the Government of India’s policy think tank), working with the IT industry body NASSCOM and Boston Consulting Group (BCG), estimates that under business as usual the tech services workforce of 7.5 to 8 million in 2023 could shrink to about 6 million by 2031. With timely, large-scale investment in AI engineering and system design skills, it sees room instead for up to 4 million new AI linked roles [5].
Demography raises the stakes. In 2022 young people made up 82.9% of the unemployed in India, and those with secondary or higher education accounted for 65.7% of the jobless young [22]. If AI trims entry level hiring in India as the US data hints, it will fall on the very cohort already queuing for a first job. The American evidence carries a second warning: much of the early strain shows up among graduates who have stopped searching, a group that headline unemployment rates, in India as in the US, do not count.
India exports the very tasks AI does best, and its jobless are already young and educated.
7. The AI Survival Scenario: What Needs Updating
Because Massenkoff and McCrory [1] set out thresholds, we can say in advance what each possible future would look like in empirical data, and check it against the numbers as they arrive.
AI Labor Market Scenarios
Tracking our current trajectory against potential future outcomes
Slow Diffusion
Exposure creeps toward capability, unemployment gap near zero
Entry Squeeze
Unemployment flat, young hiring falling, sidelined growing
White Collar Recession
Exposed unemployment doubles (3% to 6%)
Severe Displacement
Top 10% lost → exposed 43%, national 13%
Entry Squeeze — In Depth
Current status. Entry-level hiring has slowed significantly as firms substitute junior tasks with AI. While overall unemployment remains low, young and recent graduates are increasingly sidelined, taking longer to find relevant work.
Survival, in this sense, is not a single act of reskilling. It is an adjustment at three levels at once.
Surviving the AI Shift: Who Needs to Do What?
Invest where a human has to sign off
As AI makes routine competence cheap and abundant, value migrates to what models cannot complete alone: accountability, verification, negotiation, and high-stakes judgement.
Learn prompt workflows, verification loops, and metered API tools to multiply personal output speed.
Specialize in error detection, edge-case handling, and domain-specific regulatory verification.
Demonstrate tangible shipped products, code repos, and verified problem solving over paper credentials.
For the Individual: Climb Towards Judgement
The MIT economist David Autor argues that wages follow scarce expertise. Borrowing a line from a cartoon villain, he notes that “if everyone is an expert, no one is an expert” [23]. As AI makes routine competence plentiful, value migrates to what it cannot finish alone: judgement, verification, negotiation and accountability. Learn the tools, by all means, but invest most in the part of the work where a human has to sign.
For Employers and Universities: Rebuild the First Rung
First drafts, basic code and data cleaning were never glamorous, but they were how beginners became experts. Fairlie and Wu [2] put it plainly: entry level tasks produce output, but they also train people, through learning by doing and time beside experienced colleagues.
If AI absorbs those tasks, and remote work thins the mentoring around them, apprenticeship will have to be designed on purpose rather than inherited by accident. Erik Brynjolfsson has warned of the “Turing Trap” [24]—the managerial temptation to build machines that merely imitate and substitute cheap human labor, instead of extending and augmenting what people can do. Which path a firm takes is a managerial choice as much as a technical one.
For Policy and Statistics: Measure Hiring, Not Just Jobs Lost
Unemployment is a late and blunt signal, and the class of 2026 shows why. Its strain surfaced not in the headline rate but among the sidelined, the mismatched and the seasonal peaks. India would be better served by occupation level tracking of job entry in the Periodic Labour Force Survey, regular counts of graduates who want work but have stopped searching, an AI usage module in enterprise surveys, and social protection that travels with early career workers from job to job. The time to build the instruments is before the weather turns.
Epilogue: The Door, Not the Floor
Keynes, writing in 1930, blamed the pain of rapid technical change on the “growing-pains of over-rapid changes”. Nearly a century later, two careful studies agree that AI has not pulled the floor from under working people, not even from under the class of 2026. What it may be doing, more quietly, is narrowing the door through which the next generation walks in.
Which brings us back to Marienthal. If work gives us a shape to the day, company, purpose and a sense of who we are, then the deepest risk is not only lost income. It is a generation that never quite gets to begin. The numbers are early and the effects small. That is exactly why this is the moment to watch them.
AI and the Future of Work: Are Entry-Level Jobs Disappearing?
Notes
- pp (percentage points). The plain difference between two percentages. A rise from 7.3% to 7.8% is 0.5 pp, not 0.5%.
- Observed exposure. The share of the tasks in an occupation that AI is both able to do and actually used for at work, weighted by the time each task takes, with automated use counted in full and assisted use counted at half [1].
- Current Population Survey (CPS). The monthly household survey run by the US Census Bureau and the US Bureau of Labor Statistics, covering more than 130,000 people. It is the source of official US unemployment figures [2].
- Statistically significant. A change too large to be put down to chance in the sample, usually judged at the 5% level. “Not significant” does not prove there is no effect; it means the data cannot yet rule out zero.
- Sidelined. Graduates outside the labour force who say they want a job but have not actively searched in the past four weeks, and so are left out of official unemployment [2].
- Teleworkable occupations. Occupations whose tasks can plausibly be done from home, as classified by Dingel and Neiman [19].
References
- Massenkoff, M. and McCrory, P. (2026). Labor market impacts of AI: A new measure and early evidence. Anthropic, 5 March. Link
- Fairlie, R. W. and Wu, J. (2026). The early impacts of AI on employment among recent college graduates. NBER Working Paper 35796. National Bureau of Economic Research, September. Link
- Bonney, K., Breaux, C., Buffington, C., Dinlersoz, E., Foster, L., Goldschlag, N., Haltiwanger, J., Kroff, Z. and Savage, K. (2026). The microstructure of AI diffusion. Center for Economic Studies Working Paper CES-WP-26-25. US Census Bureau.
- Gmyrek, P. et al. (2025). Generative AI and jobs: A refined global index of occupational exposure. ILO Working Paper 140. International Labour Organization and NASK. Link
- NITI Aayog (2025). Roadmap for job creation in the AI economy. Frontier Tech Hub, with NASSCOM and BCG, 10 October.
- Jahoda, M. (1982). Employment and unemployment: A social-psychological analysis. Cambridge University Press.
- Gates, B. (2026). The turbulent AI era is here. The choices we make now are critical. Gates Notes, 26 August. Link
- Kinghorn, J. (2026). ‘People talk about AI reducing jobs, complete nonsense’: Nvidia’s Jensen Huang criticises economic doomerism on GTC stage. PC Gamer, 2 June. Link
- Keynes, J. M. (1930). Economic possibilities for our grandchildren. The Nation and Athenaeum, 11 and 18 October. Link
- Blinder, A. S. (2009). How many US jobs might be offshorable? World Economics, 10(2), 41.
- Ozimek, A. (2019). Overboard on offshore fears. SSRN. Link
- Massenkoff, M. (2025). How predictable is job destruction? Evidence from the Occupational Outlook. Working paper.
- Eloundou, T., Manning, S., Mishkin, P. and Rock, D. (2023). GPTs are GPTs: An early look at the labor market impact potential of large language models. arXiv:2303.10130.
- Ramp (2026). Ramp AI Index. Link
- US Bureau of Labor Statistics (2025). Employment projections 2024 to 2034.
- Gimbel, M., Kinder, M., Kendall, J. and Lee, M. (2025). Evaluating the impact of AI on the labor market: Current state of affairs. The Budget Lab at Yale.
- Brynjolfsson, E., Chandar, B. and Chen, R. (2025). Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab.
- Emanuel, N., Harrington, E. and Pallais, A. (2026). Remote work leaves younger workers sidelined. Liberty Street Economics, Federal Reserve Bank of New York.
- Dingel, J. I. and Neiman, B. (2020). How many jobs can be done at home? Journal of Public Economics, 189, 104235.
- Cazzaniga, M. et al. (2024). Gen-AI: Artificial intelligence and the future of work. IMF Staff Discussion Note SDN/2024/001. International Monetary Fund.
- World Economic Forum (2025). The Future of Jobs Report 2025. Geneva.
- Institute for Human Development and International Labour Organization (2024). India Employment Report 2024: Youth employment, education and skills. New Delhi.
- Autor, D. (2024). AI could actually help rebuild the middle class. Noema Magazine, 12 February. Also NBER Working Paper 32140. Link
- Brynjolfsson, E. (2022). The Turing Trap: The promise and peril of human-like artificial intelligence. Daedalus, 151(2), 272 to 287.
Data: observed exposure at task and job level is available from the Anthropic Economic Index. Current Population Survey microdata are published by the US Census Bureau and the US Bureau of Labor Statistics.
Acknowledgements
Stat and State acknowledges the two studies on which this dispatch is built: Labor market impacts of AI: A new measure and early evidence by Maxim Massenkoff and Peter McCrory, published by Anthropic [1], and The early impacts of AI on employment among recent college graduates by Robert W. Fairlie and Jane Wu, NBER Working Paper 35796 [2]. Their data, tables and findings are summarised and adapted here with full attribution, and any errors of interpretation are our own. NBER working papers circulate for discussion and have not been peer reviewed; readers should treat [2] as early evidence.
Declaration on the Use of AI
Stat and State used an AI writing assistant (Claude, developed by Anthropic) to improve the structure, language and readability of the text, to draft tables and figures, and to check figures and citations for consistency. Anthropic also published source [1]; that source was used on its merits and cited like every other. Every data point was checked against its original source, and the Stat and State Lead Team reviewed, edited and approved the final text. Responsibility for the content, analysis and any errors rests with Stat and State.