The First Paycheck Goes Missing

Two Findings That Disagree

Mara: Irene Chang is 21, with a degree in industrial and systems engineering. She says she applied for about 450 entry-level data-analyst jobs, landed 19 interviews and got no offer. In her words: “I feel like the entry-level skills that you have, they’re important, but also something that AI can do very efficiently.” Now set her story against two reports. In May 2026, authors at Yale's Budget Lab said their analysis so far found no statistically or economically significant AI effect on employment or real hourly wages. That September, a Census working-paper abstract reported that graduates of the most AI-exposed tenth of majors had 13% lower initial earnings, a drop its authors compared to graduating into a large recession.

Eli: Those can't both be the whole picture. Nothing significant on one side, a recession-sized dent in first paychecks on the other.

Mara: Maybe each one is a piece of it. And together they raise a plain question: if artificial intelligence comes for good jobs, whose income goes first? It matters if you're counting on a degree, paying for one, or waiting for a kid to land a first job.

Eli: The picture most people carry is a mid-career professional, years into a good job, losing the salary.

Mara: This story is about something quieter: the moment someone gets hired for the first time, or doesn't. Today's angle is whose paycheck artificial intelligence reaches first.

Eli: We follow Stanford economist Erik Brynjolfsson from a call center where an artificial intelligence assistant lifted brand-new workers, to payroll records where beginners start to go missing, and on to what that means for any plan to replace lost income.

A Tutor at the Elbow

Mara: The first clue comes from a 2023 paper Brynjolfsson co-authored, published through the National Bureau of Economic Research. Picture a customer-support agent in a live chat. Beside the conversation, an AI assistant offers suggested replies and links to documentation. Only the agent sees them, and the agent decides whether to use them fully, partly, or not at all.

Eli: How many agents are we talking about?

Mara: Five thousand one hundred seventy-nine, across three million chats.

Eli: And who got a lift?

Mara: One striking result came from the newest people. The authors reported that agents with less than a month on the job resolved 46 percent more issues per hour than agents of the same tenure without the assistant. And across many outcomes, agents two months in with AI did as well as or better than agents with more than six months of experience who didn't have it.

Eli: Two months in, performing like someone with six.

Mara: For beginners. Among the most skilled agents, the authors reported no productivity increase at all.

Eli: Which leaves an uncomfortable thought. If a beginner is suddenly as good as a veteran, does a company need as many beginners?

Mara: The authors raised it themselves. Firms might hire more novices, de-skill the positions, or build AI that replaces lower-skill workers entirely. Their data couldn't show which, they said; the effect on demand for workers was unclear.

The Canaries

Eli: Two years later, Brynjolfsson went looking at hiring. In August 2025, he and two Stanford colleagues, Bharat Chandar and Ruyu Chen, released a working paper built on records from ADP, the payroll company, covering millions of American workers. It reported a gap opening for workers ages 22 to 25 in the occupations most exposed to AI.

Mara: And his team chose a memorable name for those young workers: 'canaries in the coal mine.'

Eli: A cheerful thing to be called at 23. The bird you watch to see if it stops singing.

Mara: Their point is early warning. So they kept watching. A revised version, in August 2026, runs the payroll data through June of that year.

Eli: And what did the revision show?

Mara: Employment for 22-to-25-year-olds in the most AI-exposed occupations stood 19 percent below the path it would have followed if it had kept pace with less-exposed peers.

Eli: So not 19 percent of young people fired. Nineteen percent below a comparison line.

Mara: A shortfall against the path their peers set. And the authors wrote that the divergence had widened steadily since they first documented it in August 2025.

Eli: Where does a gap like that come from? Layoffs?

Mara: That's the part that stands out. The authors said the gap operates mainly through reduced hiring of young workers, rather than increased separations.

Eli: Not so much people pushed out as fewer people let in.

Mara: And not everywhere AI shows up. Declines were concentrated in occupations where AI mostly substitutes for tasks people do. Where it mostly complements workers, employment was flat or rising, especially for experienced workers.

Eli: And the experienced people inside the exposed jobs?

Mara: No comparable gap, the researchers reported.

Eli: That's the call-center result turned inside out. The machine helped beginners, and now beginners are the ones missing.

Mara: It's tempting to draw that line straight. The team calls these descriptive results, not causal estimates, and in their words, 'the facts we document may in part be influenced by factors other than generative AI.' Some of the divergence, they note, began before generative AI arrived.

Eli: But they're not waving it off, either.

Mara: No. Brynjolfsson's team finishes that same sentence by writing that their results are 'consistent with the hypothesis that generative AI has begun to affect entry-level employment.'

The Price of the First Rung

Eli: Behind every missing hire is someone sending out applications. Irene Chang is 21, an industrial and systems engineering graduate. She's been looking for entry-level data-analyst work.

Mara: How has it gone?

Eli: As she put it: 'I feel like the entry-level skills that you have, they're important, but also something that AI can do very efficiently.'

Mara: And she's describing the substitution side of the Stanford data in a single sentence.

Eli: The Census paper gives that experience a size. Its authors, Cody Orr, Lee Tucker and Lawrence Warren, used administrative records on college graduates and compared majors by how exposed their fields are to AI in the job market.

Mara: And for the most exposed tenth of majors?

Eli: In regression-adjusted estimates, meaning after accounting for other measurable differences, those graduates were five percentage points less likely to land initial employment. Their initial earnings, measured over a full quarter, were 13 percent lower.

Mara: A smaller chance of a first job, and a smaller first paycheck if you get one. That's the decline the authors compare to graduating into a large recession.

Eli: And the abstract says employment, earnings and job-switching patterns for those majors began to diverge right after ChatGPT arrived in late 2022.

Mara: Then there's a detail in how the lost pay breaks down. The abstract attributes roughly half the drop in starting earnings to lower pay within the industries that hire these graduates. The rest comes from a shift toward lower-wage sectors, and the examples it names are restaurants and retail.

Eli: So half is the degree buying less in the usual places, and half is graduates landing in a lower-paying corner of the economy altogether.

Mara: And it doesn't simply wash out. The abstract says the effects fade as graduates move further from entering the labor market, but remain substantial for the most exposed majors.

Eli: Fading, but not gone.

A Wide-Angle Lens

Mara: Which is how the Yale finding and these can sit side by side. The Budget Lab authors explained it themselves: their data source, the government's Current Population Survey, is best suited to broad groups and somewhat underpowered for narrow ones, like recent college graduates in their twenties. If AI's effects are limited to a narrow slice of workers, they wrote, other datasets and designs would be better suited to find them.

Eli: A wide-angle lens, with the action in one corner of the frame.

Mara: And Stanford agrees about the wide view. It reports no evidence of widespread, economy-wide job displacement, and describes the adjustment as showing up in employment rather than base pay.

Eli: So the place to look isn't the pay stub. It's the hiring list.

Who Pays for the Learning

Mara: Young people are already weighing that. In an opt-in poll run by the company SimplyWise, one respondent, aged 25 to 27, wrote: 'Since a lot of jobs are being replaced by AI, I have to consider being replaced before I spend years studying for a career I can never have.'

Eli: Then there's Kayden Evans, an 18-year-old in Mesa, Arizona. He interns at Empire Cat, which sells, rents and services heavy equipment.

Mara: What's he planning?

Eli: Reporter Megan Cerullo, who profiled him, wrote that he plans to go straight from high school into an apprenticeship there. 'I wouldn't say I am worried about AI because where I want to grow is as a field technician,' he said. And: 'AI can't go out in the field and take apart an engine.'

Mara: I admire that certainty. The snag is getting trained. A national report surveying public high school students found that a third of those who wanted skilled-trades classes couldn't take them, because their school didn't offer them or had no seats.

Eli: So lay the pieces side by side.

Mara: The paycheck going missing first isn't the veteran's. Exposed jobs do skew college-educated; the Budget Lab reports workers in them are much more likely to hold four-year degrees. But inside those jobs, Stanford reports employment for 22-to-25-year-olds 19 percent below the path it would have followed, and no comparable gap for experienced workers. Census reports graduates of the most exposed tenth of majors less likely to land a first job, and earning 13 percent less at the start.

Eli: Different studies, different groups, different measures, so you can't add them into one number. But they point at the same place.

Mara: The beginner's first paycheck: smaller, or never issued, easing over the years but still substantial for the most exposed majors. The reported pressure is showing up at the entry door.

Eli: And the jobs that held up? Where AI complements workers, employment was flat or rising. Kayden's bet is on work he says AI 'can't go out in the field' to do, if young people like him can get the training.

Mara: Which sets a test for any income plan: does it reach someone who was never hired? The French newspaper Le Monde reported that researchers at the International Monetary Fund proposed heavier taxes on capital, higher unemployment benefits and investment in training.

Eli: And writing in the personal finance magazine Kiplinger, Jennifer Schwab Wangers argues that large retraining efforts have struggled for decades to return displaced workers to their earlier earnings.

Mara: For a graduate who never got that first job, there are no earlier earnings to return to.

Eli: Nearly every career starts with the easy work, where you learn the hard parts. The payroll data show fewer beginners in exposed jobs now getting the chance. As the tech outlet IT Pro summarized the Stanford research, AI is less able to replace 'the idiosyncratic tips and tricks that accumulate with experience.' If the entry rung is where people pick those up, who becomes tomorrow's experienced worker?

Mara: If machines do the easy work, someone has to decide who pays for the learning.

Eli: There's much more to learn about AI and entry-level jobs, and Erik Brynjolfsson is worth keeping up with.

Mara: Follow the economy too. We have more episodes there, so you'll have the context when new research comes out.

Eli: Thanks for listening. I'm Eli.

Mara: And I'm Mara. Until next time.