The Broken Bottom Rung: The Entry-Level Job Market Under AI

The inversion, measured

A degree used to be a hedge against unemployment, and for the young in particular it was close to a guarantee. Recent graduates have historically posted lower unemployment than the workforce as a whole, the reward for the years of unpaid tuition and foregone wages that a credential represents. That relationship has now flipped. Across the fourth quarter of 2025 and into the second quarter of 2026, the unemployment rate among recent college graduates aged 22 to 27 ran about 5.6 to 5.7 percent, above the national rate of roughly 4.2 percent. Oxford Economics and data drawn from the New York Fed's college-labor-market series describe this as the first time in approximately 45 years of comparable data that recent graduates have faced a higher jobless rate than the country at large.

Underemployment tells the same story from a different angle. Among recent graduates who do hold jobs, about 42 percent are working in positions that do not typically require a college degree, the highest underemployment share for this cohort since 2020. And the concentration of the damage is stark: new labor-market entrants, meaning people who are unemployed and looking for their first job rather than laid off from one, account for roughly 85 percent of the total rise in US unemployment since mid-2023. Their share of the unemployed reached a 37-year high in 2025. This is not a story about the whole labor market souring. It is a story about the door in.

The New York Fed's own landing page for this research carries an access restriction that is worth naming rather than glossing over: the primary dataset was not directly reachable during this research and the figures above are drawn through Federal Reserve commentary and blog coverage of it, not the raw series itself. That is a secondary-sourcing caveat, and it is treated as one throughout this study.

For the first time in roughly 45 years of data, recent college graduates in the United States face a higher unemployment rate than the country as a whole.

The entry-level economy, by the numbers

The inversion shows up just as clearly on the demand side, in what employers are actually posting and hiring for. Indeed Hiring Lab's July 2026 read of the US posting data found entry-level job postings down about 7.5 percent year over year as of May 2026, while senior-level postings rose about 14.7 percent over the same period. Even where hiring is recovering, it is recovering upward, not downward: of the gains in software-related postings specifically, 71 percent went to senior roles, and 37 percent of the gains went to jobs with an AI-related title in them. Whatever bounce is underway in the labor market for skilled technical work, it is not reaching the bottom rung.

The picture inside individual firms is sharper still. SignalFire's 2025 State of Tech Talent report found new-graduate hiring at the largest technology companies down roughly 50 percent versus 2019. New graduates now make up about 7 percent of Big Tech hires, itself down 25 percent from 2023, and at venture-backed startups new graduates account for under 6 percent of hires. These are firms with the deepest AI adoption and the clearest incentive to substitute machine assistance for junior headcount on routine coding, testing, and documentation work, exactly the tasks a first-year engineer used to be hired to do.

None of this means hiring has stopped. It means hiring has tilted, sharply and consistently, toward people who already have a track record. The problem is structural rather than cyclical in one specific sense: a market that hires almost exclusively at the senior level cannot, by definition, be replenishing its own senior tier five years from now.

What the payroll data show

Postings and hiring surveys can be read multiple ways, so the most useful evidence here is the study designed specifically to isolate the AI channel from everything else moving in the economy at once. Researchers at Stanford's Digital Economy Lab, Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, used ADP payroll microdata covering millions of US workers and published "Canaries in the Coal Mine" in November 2025. Their central finding: workers aged 22 to 25 in occupations most exposed to generative AI experienced roughly a 16 percent relative decline in employment since the technology's widespread adoption, a decline that survives controlling for the shock each individual firm was experiencing at the time. Older and more experienced workers in the identical occupations were stable or growing over the same window.

The mechanism the paper describes matters as much as the headline number. The decline concentrates specifically where AI automates the work rather than merely assisting with it, which the authors locate in software development, customer service, and clerical occupations, the classic entry-level knowledge-work triad. And the adjustment shows up in headcount, not in pay: firms are not paying junior workers less for the same work, they are employing fewer of them. That distinction matters for how the story gets told. This is not wage compression at the bottom. It is a narrowing door.

A secondary and more loosely sourced data point points the same direction inside one occupation specifically: junior software developers aged 22 to 25 were reported down about 20 percent from their late-2022 peak by mid-2025. That figure travels through fewer independent sources than the Stanford paper and is tiered accordingly, but it sits inside the same occupation the Stanford study flags as most exposed, and it moves in the same direction.

The experience paradox

Put the demand-side and payroll evidence together and a specific, self-reinforcing loop appears, one that Aneesh Raman, LinkedIn's chief economic opportunity officer, named in a widely discussed 2025 essay: the career ladder is becoming a climbing wall. Employers still ask for experience on nearly every posting, junior or not. But the roles that historically manufactured that experience, the first analyst seat, the first junior developer post, the first paralegal desk, are exactly the roles most exposed to automation, because they are also the most standardized, most rules-governed, most documentable work in any given profession. AI systems learn fastest from the tasks that are easiest to specify, and junior work is disproportionately made of exactly those tasks.

The result is a paradox rather than a simple contraction. It is not that entry-level jobs have vanished outright, it is that the jobs which remain increasingly ask for the experience that only an entry-level job could previously supply. A new graduate cannot get the first job that would have qualified them for the second, because the first job is the one most easily assisted or replaced by a machine, and the employer posting the second job wants a candidate who has already cleared that bar. Each individual employer's decision to hire experienced people over juniors is locally rational. In aggregate, it removes the rung the whole ladder depends on.

This is also the study's clearest connection to the broader reversal this series documents. Prior waves of automation, industrial robots, early enterprise software, hit routine physical and clerical labor and largely spared white-collar, degree-holding work. Generative AI runs the opposite direction: Pew Research finds 27 percent of bachelor's-or-higher workers sit in the most AI-exposed jobs versus 12 percent of workers with only a high school education, and the same educational gradient shows up independently in OpenAI's "GPTs are GPTs" analysis, in Brookings' occupational exposure work, in the IMF's finding that roughly 60 percent of advanced-economy employment carries meaningful AI exposure, and in Anthropic's own usage logs, where 57 percent of observed tasks are augmented and 43 percent automated outright. Entry-level white-collar work sits squarely inside that reversal. It is educated, it is cognitive, and it is, disproportionately, standardized enough to automate. The broken bottom rung is what that inversion looks like from inside a single career, rather than across a whole labor market.

A new graduate cannot get the first job that would qualify them for the second, because the first job is the one a machine can most easily assist or replace.

Is it AI, or is it something older

The complication is that entry-level hiring has weakened before, for reasons that have nothing to do with generative AI, and several of those older forces are plausibly still operating today. The most direct competing explanation is a hiring hangover: technology firms over-hired sharply during 2021 and 2022, when near-zero interest rates made growth cheap and competition for talent was fierce, and much of the subsequent pullback in junior hiring may simply be correction from that overshoot rather than a new AI-driven regime. Higher interest rates since 2022 have independently raised the cost of speculative headcount of every kind, junior or senior. Tariff and trade uncertainty through 2025 and 2026 has added a further reason for employers to defer hiring decisions generally. None of these forces are about AI at all, and a market correcting from a hiring bubble should be expected to hit the newest, least-tenured workers hardest regardless of what technology is available.

The case for an AI-specific channel rests on a narrower but more precise piece of evidence: the Stanford payroll study's finding survives controlling for firm-level shocks, meaning the 22-to-25 decline in AI-exposed occupations is not simply explained by which firms were struggling for macroeconomic reasons. It also concentrates specifically in the occupations the researchers independently classify as most exposed to automation, rather than spreading evenly across the youth labor market, which a pure interest-rate or hiring-hangover story would predict less cleanly. If the deterioration were purely macroeconomic, junior workers in AI-exposed and AI-sheltered occupations should have suffered roughly alike. They have not.

Both readings can be true in different proportions, and the more defensible position is that they probably are. The base rate of the correction is very likely a real interest-rate and post-pandemic effect, and the AI-specific channel very likely explains an incremental but non-trivial share on top of it, concentrated in the specific occupations named above. What the current evidence does not support is a claim of precise attribution, an assertion that some stated percentage of the entry-level collapse is "caused by AI" as opposed to rates, trade policy, or overhiring correction. No source in this study's spine makes that claim with the confidence it would require, and this study does not manufacture one.

India: the same pattern, sharper

India's information-technology sector, the world's largest hub of exactly the standardized, English-language, rules-governed knowledge work most exposed to generative AI, shows the same inversion at greater magnitude. Annual fresher hiring across the sector fell from roughly 600,000 in FY2022 to about 120,000 in FY2025, a decline on the order of 80 percent, while entry-level tech openings were down about 44 percent year over year into 2026. This occurred as the industry's overall revenue kept climbing, the same decoupling of dollar output from junior headcount this series has documented in India's broader services economy.

The human cost shows up starkly in graduate unemployment. The ILO puts India's graduate unemployment rate at roughly 29 percent, on the order of nine times the unemployment rate among people who cannot read or write, an inversion that would have seemed absurd a generation ago, when a degree was the surest route out of informal work. The 2026 State of Working India report found that about 40 percent of graduates aged 15 to 29 who are in the labor force are in open unemployment, meaning actively looking and not finding anything.

India's labor statistics carry a well-known measurement problem worth stating plainly: the Periodic Labour Force Survey's annual and monthly releases, CMIE's independent panel data, and ILO modeled estimates use different sampling frames and definitions, and headline unemployment percentages can vary meaningfully depending on which is cited. The direction of the finding, severe and rising distress specifically among educated young jobseekers, is corroborated across all three sources. The precise percentage attached to it is methodology-dependent, and this study treats the figures above as tiered accordingly rather than as a single settled number.

Where the ladder still holds

The pullback is not universal, and treating it as a monolith would misstate the evidence. IBM has stated plans to roughly triple its US entry-level hiring in 2026, and McKinsey has projected hiring up about 12 percent in 2026 versus 2025. Both are counter-signals from firms with every incentive to say otherwise if AI made junior hiring obsolete for them specifically, and both suggest the collapse is concentrated by sector and function rather than sweeping every employer at once. Firms whose competitive advantage rests on judgment, client trust, or regulated expertise appear to be treating junior talent as a longer-horizon investment even as technology firms cut back.

The occupations most resistant to the squeeze share a common feature: they are physical, on-site, or built on human judgment that current AI cannot yet substitute for at the entry level. Skilled trades, clinical healthcare roles, direct caregiving, and mental-health work appear consistently on career-site and industry rankings of AI-resistant entry paths. Those rankings come from career-marketing sources rather than peer-reviewed labor economics and should be read as directional signals, not precise scores, but they converge with the Stanford paper's own finding that AI's effect concentrates specifically in software, customer service, and clerical work.

The collapsing side of that same list is now familiar from the evidence above: entry-level coding, tier-one customer support, data entry, and paralegal document review all sit near the top of every account of what generative AI reaches first. The UK's largest professional-services firms have made public cuts explicitly citing AI, KPMG's 2023 graduate intake fell from 1,399 to 942, a decline of about 29 percent, though this figure travels through secondary UK press coverage rather than a primary disclosure and is tiered as emerging rather than established. Taken together, the pattern is uneven rather than universal: the ladder is breaking hardest where the work was most standardized to begin with, and holding up best where it was not.

The extreme scenario, and its critics

The most alarming number attached to this debate did not come from a labor economist. In May 2025, Anthropic chief executive Dario Amodei forecast that AI could eliminate up to half of entry-level white-collar jobs within one to five years and push unemployment toward 10 to 20 percent, a projection widely reported by Axios and quoted across the press since. It is worth stating plainly what kind of claim this is: a projection from the CEO of a frontier AI lab with a direct commercial interest in the perceived scale and inevitability of AI's labor-market impact, not a peer-reviewed estimate, and it has been publicly disputed by economists who note that it rests on no disclosed model or dataset. This study treats it as a contested claim from an interested party, not as evidence, and reports it because it has shaped the public conversation this series is addressing, not because the underlying evidence in this study's spine supports a figure of that magnitude.

A more measured, if still forward-looking, baseline comes from the World Economic Forum's Future of Jobs Report 2025, which surveyed employers directly rather than projecting from first principles. It identifies data-entry clerks, bank tellers, cashiers, administrative assistants, and bookkeeping or payroll clerks among the roles employers expect to decline fastest, and AI and machine learning specialists, big-data specialists, and cybersecurity specialists among the fastest-growing by rate. This maps closely onto the occupational pattern the rest of this study's evidence shows: routine, rules-based, entry-adjacent roles shrinking, and roles requiring either technical AI fluency or judgment growing. It is still a projection, drawn from what employers say they expect rather than what has already happened, and this study marks it as such.

The gap between these two framings, a specific interested-party forecast of mass elimination and a more diffuse employer-survey projection of role-level shifts, is itself informative. The measured evidence in this study, the postings data, the payroll study, the Big Tech hiring figures, supports a real and significant deterioration already underway. It does not, on its own, support a specific forecast of how much worse it gets or how fast. Anyone reading this debate should hold that distinction: what has already happened is documented; what happens next is a scenario, stated by people with varying incentives to make it sound larger or smaller than the data alone would justify.

What has already happened to entry-level hiring is documented. What happens next is a scenario, offered by people with varying incentives to make it sound larger or smaller than the data alone would justify.

The limits of this reading

Several cautions bound this study's argument. The clearest is causal attribution: the entry-level labor market has weakened by every measure this study cites, but disentangling the AI-specific share from interest rates, tariff uncertainty, and correction from 2021-2022 overhiring is not something the current evidence can do with precision, and this study has deliberately declined to manufacture a percentage where the underlying sources do not supply one. The Stanford payroll study is the strongest single piece of evidence for a distinct AI channel because it controls for firm-level shocks, but it is one study, on a specific and relatively short post-adoption window, and its authors themselves frame it as an early signal rather than a settled verdict.

Several figures in this study's spine are secondary-sourced rather than primary. The New York Fed's own recent-graduate data page returned an access restriction during this research and is cited here through Federal Reserve commentary rather than the raw series. The KPMG UK graduate-intake figure and the junior-software-developer decline both travel through trade and career press rather than a primary disclosure. India's graduate-unemployment figures vary by methodology across the PLFS, CMIE, and ILO, and the specific percentage cited should be read as one estimate among a contested range rather than a single fixed number. Each of these is tiered accordingly rather than presented as more certain than it is.

Finally, entry-level hiring is historically the most volatile part of any labor market in any downturn, AI-driven or not, because junior workers are the cheapest and fastest headcount for an employer to cut or defer. Some portion of the deterioration this study documents would likely have occurred in a slower economy with no generative AI in it at all. What survives that caveat is narrower but, this study argues, well supported: the deterioration is real, it is concentrated specifically in the standardized, rules-governed work generative AI reaches first, and it is arriving at exactly the point in a career where a worker most needs the first rung to reach the second. The magnitude attributable to AI alone is genuinely uncertain. The shape of who is affected first is not.