42% of New Graduates Are Underemployed. AI Didn't Take Their Jobs. It Took the First Rung.
Every few weeks a new headline announces that AI has destroyed the entry-level job market. Every few weeks a new working paper says it hasn't. Both are reporting real data. The confusion comes from measuring the wrong thing.
Start with what is not in dispute. The New York Fed's tracker put underemployment among recent college graduates at 42% in the second quarter of 2026, meaning four in ten are working jobs that do not require their degree. Unemployment for the same cohort ran at 5.6%, above the national rate. Census Bureau research finds that graduates from AI-exposed majors, computer science among them, are increasingly landing in retail and food service instead of the white-collar roles they trained for.
Now the counterweight. A UCLA Anderson paper by Fairlie and Wu found that this summer's unemployment rate for bachelor's holders aged 22 to 25 was not significantly higher than in any summer since 2022. A CESifo paper reached the same conclusion: no evidence of significant AI-driven hiring displacement through summer 2026.
So which is it?
Unemployment is the wrong metric
Unemployment measures whether a person has a job. It does not measure whether they have the job that leads somewhere.
A Stanford study using ADP payroll data found entry-level employment lagging specifically in AI-exposed occupations. The CESifo authors note, fairly, that payroll data measures the supply of jobs while unemployment measures something else. Put the two together and the picture is coherent: young graduates are still finding work, but a growing share of that work is not the first step of a career. They are employed and off the ladder.
That is the actual damage, and it does not show up in the unemployment rate for years. It shows up in 2031, when companies go looking for mid-level engineers, analysts, and associates with five years of experience and discover that nobody gave anyone those five years.
What the first rung used to be
Entry-level white-collar work was never primarily about output. A first-year analyst's spreadsheet was not why the bank hired them. The work was an apprenticeship disguised as a job: low-stakes tasks that taught judgment through repetition, with a senior person reviewing and correcting.
Those low-stakes tasks are precisely what AI does well now. First-draft memos, data cleanup, basic code, research summaries, test cases. A manager who once delegated them to a junior hire now delegates them to an agent, gets the result in minutes, and does not have to train anyone.
The economics are obvious, and individually rational. Collectively, they remove the mechanism by which a profession reproduces itself. We wrote in One Company Hit $50M in Revenue With Six Employees that headcount is no longer evidence of anything. The flip side is that a six-person company has no room for anyone who is still learning.
If you are a founder
This is your problem sooner than you think, for two reasons.
You will need senior people, and there will be fewer of them. The pipeline that produces experienced operators is narrowing now. In five years, the senior engineer or controller you need will be scarcer and more expensive. The companies that train people today will have a talent advantage their competitors cannot buy.
Juniors who grew up with agents are better than you think. The best early-career hires we have seen this year are not competing with AI. They are fluent in supervising it. A 23-year-old who can run five agents, check their work, and know when something is wrong is a genuinely leveraged hire, and they are available right now because most companies stopped looking.
What we recommend in the studio:
- Hire one junior for every few seniors, deliberately. Treat it as an investment in a future senior, not a cost center.
- Redesign the apprenticeship. If the agent does the first draft, the junior's job is reviewing, correcting, and explaining what the agent got wrong. That builds judgment faster than doing the draft by hand ever did.
- Make review a teaching moment. Seniors should review the junior's review, not the agent's output. That keeps the human learning loop intact.
- Pay for potential. The salary premium for proven experience is going to grow. The discount on unproven talent is the arbitrage.
If you are the graduate
The advice that worked in 2019 — get any job at a good company and climb — is weaker now, because the climb has fewer handholds. What works instead is evidence. Build something real with the tools, ship it, and show your judgment in public. Small companies with lean teams cannot afford to train you from zero, but they will hire someone who already demonstrates they can supervise the machines.
The honest version
The data says AI has not caused mass unemployment among young graduates. That is true, and it is the less important fact. The more important fact is that the economy is quietly deleting the work that turned graduates into professionals, and nobody owns the problem of replacing it.
Companies optimized for this quarter will keep deleting it. The ones optimized for the next decade will rebuild the ladder on purpose.
