Every argument about whether AI is taking jobs eventually runs into the same wall: the top-line unemployment rate keeps looking fine. So people who want to believe nothing is happening point at the fine number, and people who want to believe everything is happening point at the layoff headlines, and nobody moves.
A team of Stanford economists led by Erik Brynjolfsson decided to stop arguing about the average and look underneath it. They pulled administrative payroll records from ADP — the largest payroll processor in the United States, covering millions of real paychecks rather than survey guesses — and split the workforce by age and by how exposed each occupation is to generative AI. The refreshed cut of that data made the rounds this week, and it is not a subtle result.
The number that should scare a 23-year-old
For workers aged 22 to 25 in the most AI-exposed occupations — software developers, customer support, junior auditors, entry-level analysts — employment is down about 13% since late 2022, which is to say since ChatGPT shipped. That decline is still steepening: the latest data has this group shrinking at roughly 4% a year, up from the 2.8% the researchers measured a year ago. The trend didn’t peak and fade. It sharpened.
Now the part that makes the average look fine. In that same young age bracket, the least-exposed jobs are growing about 2% a year. Bartenders and home health aides are doing fine. The bottom fell out specifically under the desk jobs — the ones that were supposed to be the safe, aspirational, “get a degree and sit at a computer” path.
It’s a rung, not a wave
The comforting frame for two years has been the wave: AI rises, everyone gets a little wet, we all adapt together. The ADP data describes something narrower and meaner. The damage is concentrated almost entirely at the bottom of the ladder.
Look at the age gradient. The 22-to-25s in exposed roles are down. Mid-career workers aged 31 to 34 are also contracting, but only about 1.7% year over year. And workers 35 to 40? Growing at 2%. Same exposed occupations, opposite outcomes, and the only variable that flips the sign is how many years of experience you already have.
That pattern actually tells you the mechanism. AI is good at exactly the tasks a junior person is hired to do — the first draft, the boilerplate code, the routine ticket, the reconciliation nobody enjoys. It is much worse at the judgment, context, and relationship-holding that a senior person carries. So companies aren’t firing their experienced staff and replacing them with a model. They’re quietly not hiring the next cohort of juniors, because a mid-level employee plus a chatbot now covers the work that used to justify three trainees. The senior keeps their job. The 23-year-old never gets theirs.
The LostJobs read
The reason this study matters more than another round of layoff headlines is that it decouples “AI is affecting employment” from “the unemployment rate went up.” Both can be true. If you stop hiring entry-level workers, the people already employed stay employed and the headline stat barely twitches — the pain shows up as a door that quietly stops opening for the people trying to get in. That is the least visible kind of economic damage and the hardest to organize against, because the victims are diffuse, young, and not yet inside any building to walk out of.
If you’re early in your career, the actionable read isn’t “learn to prompt.” It’s that the bottom rung is the exposed rung, and the fastest way off it is whatever gets you to the judgment layer — the work that requires holding context a model can’t, owning a relationship a model can’t, or making a call someone will actually blame you for. That was always the goal. The ADP data just shrank the amount of time you have to reach it, and removed the training-wheel jobs that used to carry you there.
Sources: Time, Stanford SIEPR, Fortune, CBS News.
