The August jobs report was good. Per the Bureau of Labor Statistics, employers added 162,000 jobs, far above forecasts, and unemployment held at 4.1%.

Now look at who that number leaves out. Per the New York Fed, unemployment for recent college graduates, ages 22 to 27, was about 5.6% in the second quarter. Underemployment, meaning grads working jobs that don't need a degree, edged up to 42%.

The headline rate is 4.1%. The rate for people who just did everything they were told to do is higher.

The strongest evidence for an AI effect comes from payroll records, not surveys.

Per a Stanford Digital Economy Lab update published August 12 by Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, using ADP payroll data: between November 2022 and June 2026, employment of workers aged 22 to 25 in the most AI-exposed jobs fell about 11%. Workers the same age in the least-exposed jobs grew about 10%.

Their summary: young workers in highly exposed jobs are now "about 19% below where it would be" if they'd kept pace with peers in less-exposed work.

The split matters. The declines are "concentrated in occupations where AI usage tends to automate human tasks." Where AI is used to help workers rather than replace tasks, "employment is flat or rising, particularly among more experienced workers."

The Dallas Fed found the same pattern, and then found something more interesting: how it happens.

Per Tyler Atkinson and Shane Yamco at the Dallas Fed, workers 22 to 25 in the most exposed occupations saw a 13% employment decline between late 2022 and September 2025. The drop came mainly from "a decline in people transitioning directly from out of the workforce into employment." Not layoffs. Not people losing jobs they already had.

Read that twice, because it changes the whole story. The headline version of AI and jobs is a pink slip. The data version is a door that stays shut. Experienced people keep their seats. The first rung of the ladder is what's disappearing.

That's also why it hides so well in a strong jobs report. A layoff shows up. A job that was never posted doesn't.

If you're a student or a recent grad: "entry-level" now often means "already has experience." Internships, apprenticeships, freelance work and anything with a real output you can show are worth more than they were five years ago. The Stanford data says jobs where AI helps people are holding up better than jobs where it replaces tasks, so aim for roles where you'd be using the tool, not competing with it.

If you're a parent paying for college: ask the school one direct question: what share of last year's grads in this major were employed in a job that needed the degree within six months? If they can't answer, that's an answer.

If you manage people: the entry-level work AI is absorbing is also how you used to train your future mid-level staff. Cutting the junior hire saves money this year. Count what it costs you in 2031 when there's nobody ready to promote.

If you're mid-career: the same data shows experience is currently the protection. That lasts as long as your experience includes knowing how to use the tools.

Forest ecologists and fisheries scientists have a name for this: recruitment failure.

A forest can look completely healthy. Big trees, full canopy, nothing dying. But if you walk the floor and there are no saplings, because deer eat every seedling or the shade is too heavy, that forest is already in decline. You just can't see it yet, because mature trees live for decades. By the time the old trees start falling, there's a 30-year gap with nothing to replace them.

Fisheries learned it the hard way. Counts of adult fish can hold steady for years while the number of young fish reaching adulthood quietly collapses. By the time the adult count drops, the collapse is locked in.

The labor market is showing the same signature. The adult population, meaning experienced workers, is stable or growing. The overall numbers look fine. What's thinning is recruitment: people entering the workforce into the jobs that used to train them. A jobs report counts the canopy. It doesn't count saplings.

What to watch from here: the New York Fed's next recent-graduate update in November, and whether entry-level hiring shows up in company earnings calls as a line item anyone is proud of cutting.

The researchers themselves won't call this proof. Stanford calls its results "descriptive patterns, not causal estimates," says the gaps shrink when you account for education, and notes some of the divergence started before ChatGPT existed. The Dallas Fed says the pattern "may not be causal" and that the overall impact so far is "small and subtle."

The Yale Budget Lab goes further. In its January 28 update, it found that measures of AI exposure, automation and augmentation "show no sign of being related to changes in employment or unemployment," and that the faster shift in the job mix "predates the widespread introduction of AI in the workforce." Its data runs through December 2025, older than Stanford's, but the point stands: a slower hiring market after the pandemic boom could explain much of what young grads are facing without any help from AI.

Claim: The New York Fed's next update of its recent college graduate data (2026 Q3, expected November 2026) will show the recent-graduate unemployment rate at or above 5.3%, and above the overall unemployment rate for all workers.

Stated confidence: 75% · Verification date: November 30, 2026 · Status: OPEN

This newsletter is the free version of what I do: take a number everyone is repeating, find out whether it holds, and show the work.

If there's a claim you keep seeing about AI and jobs, in your industry or at your company, reply with it. I'll tell you whether the source supports it and what I couldn't verify.

cokas.io — no sales call, ever.

Every prediction in The Pattern Brief carries a verification date and is revisited in a future edition — including this one.