The most quietly alarming number in the labor market right now is not a stock index or an inflation print. It is this: the unemployment rate for recent college graduates reached 5.6% in the second quarter of 2026, comfortably above the roughly 4.2% rate for the workforce as a whole. New graduates are now more likely to be out of work than the average American worker — and, remarkably, more likely than many people their age who never enrolled in college at all. The Federal Reserve Bank of New York, which tracks this every quarter, also pegged recent-grad underemployment at 42% in its August 6 release.
The instinct is to blame AI and move on. That instinct is half right, and the half that's wrong matters — because it changes which major you should pick. The damage in this cycle is landing on the bottom rung of the ladder: the entry-level job. And the entry-level job is precisely the thing that conventional "best major" advice treats as a rounding error.
The squeeze was here before AI showed up
Start with what AI did not do. Recent graduates have had a higher unemployment rate than the overall workforce since 2021 — a reversal of a decades-long pattern, and one that predates ChatGPT's public debut by more than a year. Whatever opened that gap, it wasn't generative AI, because generative AI wasn't in the labor market yet.
The New York Fed's own economists have argued that a large share of the early-career slump traces to something far more prosaic: remote work. In a June 2026 analysis, they estimated that the shift to remote and hybrid arrangements could account for roughly 64% of the rise in unemployment for young college graduates between the 2017–2019 and 2022–2024 periods, as senior workers who might have retired instead stayed put in newly flexible roles. Oxford Economics, surveying the same terrain in April, found employers trimming entry-level recruitment but concluded it was not yet seeing "meaningful job losses due to AI" at scale, attributing much of the softness to a rising supply of workers chasing a fixed number of openings.
So the honest baseline is this: the market for a first job was already tight for reasons that have nothing to do with language models. That is not a reason to ignore AI. It is the reason AI matters more, not less — because a risk multiplier is most dangerous when it lands on a system already under strain.
What's new is where the damage lands
The genuinely new signal arrived on August 12, when researchers at Stanford's Digital Economy Lab updated their study "Canaries in the Coal Mine" (Brynjolfsson, Chandar, and Chen). Their finding is specific and hard to wave away: employment for workers aged 22 to 25 in the most AI-exposed occupations is now running about 19% below where it would be if it had tracked employment for similarly aged workers in less-exposed jobs. The divergence has widened steadily since August 2025.
The mechanism is the part worth memorizing. The gap opened up through reduced hiring, not through layoffs of people already employed. Companies aren't firing their 24-year-olds in AI-exposed roles; they're quietly declining to hire the next class of them. That is the entry-level squeeze in one sentence: the tasks that make up a junior role — summarizing, drafting, first-pass analysis, routine code — are the tasks a capable model does cheapest, so the job that used to absorb new graduates is the first one a firm learns to do without.
AI, in other words, is not the whole story of the youth labor market. But it appears to be the part of the story that decides who gets hit. Remote work and worker supply tightened the aggregate. AI is redistributing the pain across fields — and it is doing it at the entry point.
Why this breaks the usual major math
Nearly every "highest-paying majors" list you'll find, including the ones from the big ranking sites, is built on one number: the median earnings of people who already work in the field. That number quietly assumes the hard part — getting onto the ladder in the first place — is a solved problem. For decades that was a fair assumption. A soft entry year was a cyclical nuisance you waited out.
When the bottom rung is the rung being pulled away, that assumption inverts. The relevant question stops being "what does the median petroleum engineer earn?" and becomes "what is the probability I land the petroleum-engineering job at all, in a market where the junior version of that job is the one being automated?" A major with a gorgeous median and a collapsing entry pipeline is not a good bet; it's a lottery ticket with a great jackpot and terrible odds. Average earnings hide exactly the risk that now matters most. (We made a related argument about how a financial-aid award letter prices the degree but hides the return; this is the same blind spot, one rung lower.)
How our model already prices the entry shock
This is the risk the DegreeOutlook AI Impact Model was built to measure, and it's worth being precise about how. Instead of scoring a degree by its median salary, the model computes a ten-year expected value of earnings for every program, as a probability-weighted blend of two futures:
EV = Pfield × (field earnings) + (1 − Pfield) × (fallback earnings)
The crucial design choice: AI exposure does not cut the salary. It cuts Pfield — the probability you actually work in your field rather than falling back to a job that didn't require the degree. That fallback is anchored at about $45,000, the New York Fed's median wage for recent grads in non-college jobs. The more a major's target occupations look like tasks a model can do, the more the model marks down Pfield, and the more weight the $45,000 fallback carries.
Three scenarios turn a severity dial on that exposure — optimistic, base, and pessimistic. And the pessimistic scenario does one thing that maps almost eerily onto the Stanford finding: it sets the fallback earnings for year one to zero. The assumption is that a graduate squeezed out at the entry point earns nothing in the field, and hasn't yet landed even the fallback job — the frictional cost of a hiring freeze that hits you before your career starts. The model, built months before this data, front-loads its penalty at exactly the moment the Stanford data now shows the real damage occurring: at hiring, not at separation.
That gives you a clean, decision-grade number. The spread between a program's optimistic and pessimistic expected value is a direct read on displacement risk — and because the model front-loads that risk into the entry years, it is precisely a measure of how much of a degree's value depends on the field job actually being there for you when you graduate. A narrow spread means the payoff holds up even if the entry market stays ugly. A wide spread means the glossy average is riding on a hiring pipeline that AI is actively thinning. You can see how each field scores on our full methodology page.
Read a major by its spread, not its average
Once you look at degrees this way, the map redraws itself. The programs that top our most AI-resistant list are not the ones with the highest medians — they're the ones whose work is hard to hand to a model. The list is dominated by hands-on and clinical programs: heavy-equipment and energy-systems technology, environmental technology, clinical laboratory science, and the applied health fields. What they share is task content that is physical, in-person, or high-touch — work that still requires a body in a room, a licensed judgment call, or a hand on a tool.
The high-spread end — the degrees on our highest AI-risk list — skews toward screen-based knowledge work whose junior tier is the most automatable slice: routine drafting, first-pass research, entry-level analysis. These are often not low-paying fields at the median. That's exactly the trap. The median looks fine; the entry pipeline is where the risk lives.
| What lowers a major's first-job risk | Why AI can't cheaply take the entry job |
|---|---|
| Physical / on-site work (equipment, energy, environmental tech) | The task needs a body at a location, not a token stream |
| Clinical / licensed judgment (nursing, clinical lab, allied health) | Regulated sign-off and hands-on care resist automation, and licensure limits labor supply |
| High-touch interpersonal work (care, skilled services) | Trust and presence are the product, not a byproduct |
| Broad screen-based entry roles (generalist analysis, routine drafting) | Higher risk — the junior tier is the cheapest for a model to absorb |
A caution on mechanism, because it's easy to draw the wrong lesson: the thing protecting nursing or clinical lab science is not the diploma or the license by itself. It's that the underlying work is physical and interpersonal, which is why the AI-exposure score is low. Licensure adds a second, separate moat — it caps how fast the labor supply can grow — but it's the task content that keeps the model from doing the job. Get those two mechanisms confused and you'll talk yourself into a credential that gates entry without protecting the work.
This is risk management, not a prophecy
None of this says AI is about to hollow out white-collar work, and the model doesn't claim it. Remember the counterweights: the NY Fed still thinks remote work is the bigger driver of the aggregate slump, and Oxford still isn't seeing AI-driven job losses at scale. A wide optimistic-to-pessimistic spread is not a verdict that a major is doomed. It's a price — a signal that the median overstates the reliable return, and that you should demand a correspondingly higher payoff (or a clearer path into the field) before you accept it.
The practical move is small and worth doing before you commit four years and any borrowed money. Pull up the major you're considering. Look past the headline salary to two things: how AI-exposed its target occupations are, and how wide the gap runs between the good scenario and the bad one. A degree whose value survives a brutal entry market is worth more today than one whose brochure number depends on a hiring pipeline that is, right now, visibly narrowing. For a fuller picture of the two forces at work here, our breakdown of the two risks every major carries and our list of AI-proof careers both start from the same data.
The ladder still leads up. It's the first rung that got harder to reach — so pick the degree that puts you on a ladder AI can't quietly saw off at the bottom.
Frequently asked questions
Does a wide optimistic-to-pessimistic spread mean I shouldn't choose that major?
No. A wide spread means the average earnings overstate the reliable return, because a chunk of that value depends on landing the field job in a market that's squeezing entry-level hiring. If the field job pays enough to justify the risk — and you have a credible path into it — the spread is a price you're choosing to pay, not a veto. It just shouldn't be invisible, which is what a median-only ranking makes it.
If remote work explains most of the entry-level slump, why look at AI exposure at all?
Because they answer different questions. Remote work and worker supply help explain how tight the overall market is. AI exposure explains which fields feel that tightness most. In a soft entry market, a high-AI-exposure field is the compounding case: the aggregate squeeze and the field-specific squeeze stack on top of each other, and the youngest workers absorb both.
Isn't this just the old "pick a practical major" advice?
It's narrower and more testable. The protective trait isn't "practical" — it's task content a model can't cheaply reproduce: physical, on-site, clinical, or high-touch work. That cuts across fields the old advice would separate (applied technical programs and licensed health degrees) and, more importantly, it flags the newly risky category the old advice missed: well-paid, screen-based knowledge work whose entry-level tier is the first thing AI learns to do.