The Entry-Level Puzzle

Workers aged 20 to 24 are unemployed at nearly twice the national rate, three credible studies blame three different causes, and employers cite AI whichever cause holds. The stock data tells us the AI-exposed occupations kept growing, and that firms stopped hiring juniors into them.

Read the companion deck (PDF)→

THE CLAIM

Since Issue 01 we have argued that the AI shock reaches the economy through prosaic channels, cost, complements, and the slow diffusion of a capability, and that in the labor market it moves through hiring before it ever shows in headcount. The entry-level data is where that argument now gets tested, and it broadly holds. AI-exposed occupations are still growing, up 10.4% in employment since 2019, while the constraint that has bound is the flow of new hires, which has sunk to a rate last seen climbing out of the financial crisis and falls hardest on the people who need a first job. The debate over whether AI or remote work broke the ladder is unresolved, and an operator can move on the one fact every firm confirms in its own numbers, which is that it has stopped hiring juniors.

The unemployment gap between ages 20 to 24 and all workers widened from 3.0 points in 2023 to 4.0 points in 2025, and the worst month read 9.2% against 4.3% for the workforce as a whole (BLS).

The JOLTS hires rate printed 3.1 in February 2026, a level the series has otherwise touched only between 2008 and 2013, when unemployment was near 10%; it arrived this time with unemployment at 4.4% (JOLTS).

The three studies that dominate the debate reach three different verdicts. Stanford finds a 16% relative employment decline for AI-exposed 22-to-25-year-olds, the NY Fed traces 64% of the rise in graduate unemployment to remote work, and Yale finds the occupational mix sitting on its pre-2022 trend.

Our own cut of the stock data cuts against the destruction story. The most AI-exposed occupations added employment (+10.4% against +3.8%) and real wages (+3.9% against +2.6%) from May 2019 to May 2025 (Second Order calculations on BLS OEWS).

Employers named AI in 40% of May's announced cuts, up from 7% in January, though the figure measures attribution rather than cause (Challenger).

01 · THE JOB MARKET

The market is safe for incumbents and shut to entrants

Unemployment for workers aged 20 to 24 averaged 8.3% in 2025 against 4.3% for the workforce as a whole, a 4.0-point gap that stood at 3.0 points as recently as 2023, and the worst months of the year printed 9.2%. The divergence itself is not in dispute. The NY Fed put unemployment for recent college graduates at about 5.7% at the end of 2025 and underemployment at 42.5%, the highest since 2020. This gap opened while the overall rate sat in the low fours, which rules out a recession as the cause and locates the strain in the entry tier (Exhibit 1).

The JOLTS hires rate printed 3.1 in February 2026 and 3.2 in April. Before this cycle the series reached those readings only between 2008 and 2013, when unemployment was climbing toward 10%, and it has arrived this time with unemployment at 4.4%. Layoffs stayed low through the same months, so firms are holding the staff they have while adding few, and the headline unemployment rate, which counts the employed and the actively unemployed but not the vacancy that was never posted, registers almost none of it. Indeed's Laura Ullrich names the visible symptom “experience creep,” the practice of asking for years of experience to fill the roles that once supplied it. Hiring is the margin firms cut first because the cost shows up slowly, since a closed entry market costs little this year and the bill comes due in about three years, when those hires would have become the mid-level staff the same firms then need (Exhibit 2).

02 · THREE STUDIES

Three credible studies name three different culprits

Stanford's Digital Economy Lab, working from ADP payroll records, finds workers aged 22 to 25 in the most AI-exposed occupations down 16% in relative employment since late 2022, while older workers in the same occupations held steady and the losses concentrated where AI automates the work instead of augmenting it. The NY Fed's June study points to a different cause, that 64% of the rise in young-graduate unemployment traces to remotable occupations, the age gap survives controlling for AI exposure, and unemployment for experienced graduates fell even as the young rate climbed toward 5.6%. Yale's Budget Lab, which tracks the aggregate occupational mix month by month, finds it moving no faster than its pre-2022 trend, 33 months after ChatGPT's release (Exhibit 3).

Stanford, the NY Fed, and Yale read as a contradiction only if each measures the same object, and they measure three different ones. Stanford measures a flow, the employment of the youngest workers inside exposed niches. The NY Fed identifies a mechanism, apprenticeship breaking down when junior staff lose proximity to the senior colleagues they would have learned from. Yale measures a stock, the standing mix of occupations, in which entry-level pain does not yet register. All three hold at once if the squeeze sits in entry hiring rather than in the stock of jobs, and they agree on that point even though none set out to make it, because none finds occupations disappearing.

03 · OUR NUMBERS

Exposed occupations grew while hiring into them froze

Using BLS occupational employment data for May 2019 and May 2025, with the exposure groups taken from Pew's 2023 classification, we measured the stock directly. The most AI-exposed occupations grew employment 10.4% against 3.8% for everything else, and real wages 3.9% against 2.6%. One sensitivity cut is worth surfacing because it changes the answer materially, since dropping office and administrative support from the exposed group lifts employment growth to 25.2% while real mean wages fall 2.3%. Firms are adding analytical white-collar roles faster than they are paying up for them, so headcount climbs in the exposed occupations even as the real wage compresses (Exhibit 4).

Washington Monthly's tally records 7% more software developers than in 2022, 10% more radiologists, and 21% more paralegals, the occupations most often named as AI casualties, and independent counts point the same way. Three facts point the same way: the exposed occupations are growing, hiring is frozen at the rates of the financial-crisis recovery, and young workers are diverging from everyone else. The driver is a belief running ahead of a measured capability, since managers now staff on the assumption that one senior worker with AI does the work that a senior and two juniors used to do. Whether that assumption holds is almost beside the point for this year's hiring, because firms are already staffing as though it does.

PwC's 2026 Global AI Jobs Barometer, drawn from more than a billion job ads, reads the same way. Early-career postings have flatlined in the most AI-exposed sectors, the “seniorised” entry roles that ask for senior skills have grown 35% since 2019, and AI-exposed junior roles are seven times more likely than the least-exposed to demand leadership and strategic thinking. The entry rung has been raised more than it has been removed, and the experience creep visible in the tape shows up, on PwC's reading, in the ad data as well.

04 · ATTRIBUTION

Employers name AI whichever cause holds

Challenger's May report makes AI the number one stated reason for US job cuts for the third consecutive month, at 38,579 cuts, 40% of the total and up from 7% in January, with year-to-date AI-attributed cuts already above the whole of 2025. The figure measures what employers say rather than what their own data shows, and the gap matters because HBR's Davenport and Srinivasan find the cuts to be anticipatory, made ahead of any demonstrated capability because investors reward the story. This is Issue 02's capital swap under another name, the same firms whose CFO headcount-growth expectations fell from 6% to 2% even as three quarters of them raised technology budgets. When the token bill lands, the graduate hire is the easiest line on the page to cut (Exhibit 5).

SECOND-ORDER EFFECT · THE SEED-CORN PROBLEM

Junior roles are the mechanism by which mid-level talent is produced, so each firm's decision to stop hiring juniors is rational on its own terms and damaging in aggregate. The juniors a firm declines to hire in 2026 are the mid-level staff it finds missing in 2029, when the same firms compete to hire that cohort back. Skipping a year of entry hiring books a saving now and manufactures a shortage later, the seed-corn trade run across a whole labor market.

The NY Fed result is why the problem is fixable at the level of a single firm. If the ladder broke where proximity ended, a structured apprenticeship built to work in a hybrid setting becomes an operating-model decision rather than a perk. Firms that rebuild it now hire into a buyer's market, with 42.5% of recent graduates underemployed, while those that wait will later pay up to reprice scarce mid-level staff. Few firms are yet acting on it, which is why those trained graduates remain cheap.

WHAT WE'D DO · THE OPERATOR'S PLAYBOOK

1. Set a graduate-intake floor. Fix entry hiring as a ratio of headcount rather than as a discretionary line, because Issue 02 showed compute budgets crowding out hiring plans, and without a floor the graduate requisition loses to the token budget every quarter it is contested.

2. Rebuild apprenticeship for how the firm now works. Named mentors, structured review cycles, and deliberate proximity, in a hybrid setting or on site, are what the 64% finding points to as the place the ladder broke.

3. Pair juniors with AI and judge them on output. If a junior with AI delivers most of what a mid-level delivered in 2023, the entry-level wage is the best unit of capacity an operator can buy this cycle.

4. Recruit from the freeze while it lasts. Underemployment at 42.5% means trained graduates are waiting tables, and buyer's markets in talent do not last; this one closes when the first large employers work out point 3.

5. Audit your own attribution before you publish it. Before naming AI in a freeze or a layoff, confirm that the firm's own numbers show it, because boards are learning to ask the question and yours should ask it first.

WATCHLIST · WHAT WOULD CHANGE OUR MIND

July 2, June jobs report. A 20-to-24 rate back under 7% for two consecutive prints would mark the divergence as peaking; a rate above 8.5% would mark it as compounding.

July 7, May JOLTS. A hires rate under 3.1 would be the lowest reading since 2013 and would harden the freeze call; a rate at 3.6 or above would begin to break it.

Early July, Challenger June. Whether AI attribution holds above 40%, and whether any firm begins walking its attribution back as the scrutiny rises.

NY Fed college labor market update, quarterly. Recent-graduate underemployment holding above 42.5% would say the door is still shut.

Stanford canaries refresh. If the 16% decline spreads from ages 22 to 25 up into 26 to 30, the AI story strengthens against our framing, and we will say so.

Falsifier. The claim that the squeeze sits in entry flows and not in occupational destruction dies if the next OEWS vintage shows the most AI-exposed occupations turning down from their +10.4% instead of extending it, with the Stanford decline broadening up the age range while the hires rate stays frozen. That combination would mean the stock is shrinking after all, and the diagnosis in this issue would be wrong.

SCORECARD · CONSENSUS VS. THE TAPE

“AI is destroying entry-level jobs.” Contested at best on the evidence we have, since 64% of the rise in graduate unemployment traces to remote work and the exposed occupations grew. AI is the headline, and the mechanism behind it is less settled than the headline implies.

“AI-exposed occupations are shrinking.” Absent from the stock data, which shows +10.4% since 2019 and more software developers than in 2022.

“Low unemployment means a healthy market.” The hires rate argues otherwise. A market can be safe for incumbents and shut to entrants at the same time, and this one is.

“AI is hollowing out the bottom of the ladder.” Half right. On PwC's reading of more than a billion job ads, early-career postings have flatlined in the most AI-exposed sectors while the “seniorised” entry roles grew 35% since 2019. The bar for entry work rose, and the work itself is still there.

NEXT ISSUE · THE CAPEX ECONOMY

AI-related investment accounted for roughly three quarters of Q1 2026 GDP growth, and without it the economy was about flat. Next week we take up what a one-trade economy does to a demand forecast, and why a planning assumption set today inherits a concentration risk no operator chose.

SELECTED SOURCES

BLS via FRED, series LNS14000036, UNRATE, JTSHIR (retrieved Jun 12, 2026) · BLS OEWS May 2019 / May 2025 with Pew Research (2023) exposure groups; Second Order calculations · Brynjolfsson, Chandar & Chen, “Canaries in the Coal Mine,” Stanford Digital Economy Lab (rev. Feb 2026) · Emanuel, Harrington & Pallais, “Remote Work Leaves Younger Workers Sidelined,” NY Fed Liberty Street Economics (Jun 1, 2026) · The Budget Lab at Yale, AI and the labor market tracker · NY Fed, The Labor Market for Recent College Graduates · Challenger, Gray & Christmas May 2026 report (Jun 4, 2026) · Davenport & Srinivasan, HBR (Jan 2026) · NPR (Jun 1, 2026) · Washington Monthly (May 29, 2026) · Indeed Hiring Lab commentary (L. Ullrich) · PwC, 2026 Global AI Jobs Barometer (Jun 15, 2026; analysis of over a billion job ads).

All chart series computed from filed primary data except Exhibits 3 and 5, which plot reported point estimates from the cited studies. Second Order is an independent research briefing, provided for discussion purposes only; it is not investment, legal, tax, or accounting advice. Copyright 2026.