The Gen Z jobpocalypse is hiding in paychecks.

In early 2026, as companies debated the effects of AI for business, BlackRock CEO Larry Fink made a prediction that became the standard-bearer for fears about AI job loss: “When this year’s college graduates enter the workforce, we could see the highest unemployment rate among them in years, even without a recession.” The summer of 2026, the season when graduates flood the job market in force, just gave its verdict. The unemployment rate for 22- to 25-year-olds with bachelor’s degrees averaged 7.3 percent from June through August, according to a working paper by UCLA Anderson’s Robert Fairlie and Jane Wu. That is below the 7.8 percent of summer 2024 and, the authors wrote, “not unusually high relative to earlier summers.”

On September 10, US Census Bureau economists published a study of administrative records covering roughly 29 percent of all bachelor’s degrees awarded between 2016 and 2024. It found that the most AI-exposed tenth of college majors saw their likelihood of initial employment fall by five percentage points after ChatGPT’s debut, and their initial full-quarter earnings fall by 13 percent, a hit the authors describe as “comparable in magnitude to the earnings losses associated with graduating into a large recession.” Two weeks later, the Federal Reserve Bank of Dallas published its own study of Texas graduates and found the same shape: exposed majors took longer to find work, and the ones who did were paid about 5 percent less.

The harm shows up as a smaller number on an offer letter, a longer gap before the first offer arrives, and a first job in a restaurant instead of an office.

AI is demoting new grads

To be unemployed, you must be without work and actively looking.

First, a computer science graduate who takes a retail job is employed. The Census study found that roughly half of the 13 percent earnings decline comes from graduates shifting into lower-wage industries such as restaurants and retail. Every one of those downgrades counts as a job filled. The New York Fed’s separate tracker puts graduate underemployment at 42 percent as of the second quarter of 2026, meaning more than two in five working graduates hold jobs that typically do not require a degree.

Second, a graduate who gives up looking vanishes from the rate entirely. Fairlie and Wu built a broader measure they call “sidelined unemployment,” which adds graduates who say they want a job but have not searched in the past four weeks. That rate hit 10.4 percent this summer, the highest of the five summers they studied, though the increase was not statistically significant.

Third, and most important, AI is declining to hire graduates. Companies freeze the bottom rung of the ladder rather than cutting the incumbents above it, so the pain lands on people who have no job to lose. A hiring freeze leaves almost no fingerprint on an unemployment rate. It shows up in time-to-first-job, in starting pay, and in the industry on the first paycheck.

The unemployment line everyone watches barely moved

Look at the chart every journalist and economist has been staring at. Graduate unemployment ran below the all-worker rate from 2016 through 2019, edged above it around 2022, and has drifted upward since. The drift is gentle: a gap averaging about 1.5 percentage points in the first half of 2026. Nothing in this line screams apocalypse, which is why so many economists have spent the summer telling anxious graduates that the data looks fine.

Now look at what happens when researchers swap the meter. Instead of asking graduates whether they are unemployed, the Census team followed actual earnings and industry records for every degree cohort, and sorted college majors by how exposed their typical jobs are to AI. The divergence appears immediately after ChatGPT’s November 2022 release, and it appears in the two places the unemployment rate cannot see: the size of the first paycheck, and the industry that signs it.

The studies and their blind spots

Each one is defensible. Each one is also blind to part of the picture, and the blind spots explain why the same labor market looks fine in one dataset and recession-grade in another.

Fairlie and Wu tested an alternative explanation and found it uncomfortably close to the AI story: unemployment was also relatively higher among graduates in occupations that can be done remotely, and employers may be reluctant to mentor inexperienced workers over Zoom. The trouble is that remote-friendly jobs and AI-exposed jobs heavily overlap, so the two explanations cannot yet be cleanly separated. Expect that argument to run through every study from here on.

The bottom rung of the ladder got deleted

The studies describe what happened, and employers describe why. The pattern in interviews with hiring firms is consistent: AI now does the research, the first drafts, the basic code, the data cleanup, which is the grunt work that entry-level hires were historically paid to do and learned from.

There is a second-order problem here that employers themselves are starting to name: the grunt work was also the training. Cheney Hamilton of research firm Bloor told CNBC that AI “raises the floor of what a junior can produce, but it doesn’t give them the judgement to know when the output is wrong.”

The escape routes, and whether they work

Students are already reacting, and the Dallas Fed data shows two of the three available exits are weak.

Exit one: graduate school. For every 10 points of AI exposure in a major’s typical jobs, 2024 Texas graduates were 1.4 percentage points more likely to enroll in a graduate program. Most stayed in the same field: two-thirds of computer science graduates who continued their studies enrolled in computer science programs. The researchers checked what that buys: “the earnings records of recent master’s degree recipients suggest there may be limited returns to formal upskilling in AI-exposed fields.”

Exit two: switch fields. Enrollment in Texas’s most exposed majors, computer science and computer engineering chief among them, fell 4.8 percent per 10 points of exposure between fall 2024 and fall 2025, while students drifted toward less-exposed fields such as nursing, education, and psychology. That is rational, but it is also a four-year bet on what AI will and will not do by 2030, made with incomplete information.

Exit three: outcompete the machine at the entry level anyway. This is the “hang in there” path that Harvard’s David Deming offered NPR’s 500-application graduates. The Census data gives it partial support, since its earnings penalty “attenuates as graduates move further from labor market entry,” though it remains substantial for the most exposed majors. The catch is timing, because starting-pay research has long shown that the first job anchors a career, and Census itself compares the current hit to graduating into a large recession. Waiting out a recession-grade start with a strategy of more applications is the approach the 450-application resumes were already trying.

Where the squeeze goes from here

The broader machinery of AI-driven cuts keeps running even as the graduate story unfolds. Through September, US employers had announced 573,195 job cuts this year, down 39 percent from 2025, yet roughly 120,000 of them, about 21 percent, were explicitly attributed to AI, the first year AI has topped Challenger, Gray & Christmas’s list of stated layoff reasons. The entry-level squeeze is the leading edge of the same shift.

A few indicators show whether the squeeze is becoming a crush. The Dallas Fed authors nominate layoffs and wage growth for prime-age workers in automation-exposed occupations, because once the damage jumps from new hires to incumbents, it is no longer hidden. Fairlie and Wu nominate the class of 2027 and beyond, since their summer test covered a market where AI adoption was still ramping. The third is the sidelined rate; if graduates keep leaving the search, that 10.4 percent will climb before any official rate does. The fourth is international. IMF chief Kristalina Georgieva called AI a “slow-moving tsunami” in Bangkok on October 13 and estimated 40 percent of jobs worldwide will be augmented, downgraded, or eliminated; Singapore’s president noted his country’s full-time graduate placement rate already fell from 79.4 to 74.4 percent in a single year.

Georgieva’s line about graduates was the one that traveled: “I don’t think any young person wants to be a barista in a coffee shop as their professional future.” The data from the Census Bureau suggests the question is no longer hypothetical. For the most exposed majors, many graduates in each recent class are already taking service jobs, at about 13 percent less than the class before them earned on the way down. The jobpocalypse everyone spent three years watching for never materialized as an unemployment spike. It arrived in smaller offer letters and longer waits for a first job.