If you graduated into a white-collar field in the last two years, you already know the number in your bones, even if you have never seen it written down. So here it is written down. According to a widely cited Stanford study drawing on payroll records from ADP, the largest payroll processor in the United States, early-career workers aged 22 to 25 in the most AI-exposed occupations have seen a roughly 13 percent relative decline in employment since generative AI went mainstream. In the same period, employment for older workers in those same jobs kept growing. The bottom rung of the ladder is thinning while the higher rungs hold.
That finding, from economists Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, landed like a thunderclap precisely because it seemed to confirm the thing everyone already feared: that AI would come for the entry-level knowledge worker first. But the more interesting story is not the headline. It is the fight that broke out immediately afterward over whether AI is actually the culprit at all — a fight that is still unresolved, and that matters enormously for anyone deciding what to study, who to hire, or how worried to be.
What the data actually shows
Start with what is not really in dispute. Something happened to young white-collar workers around late 2022, and it did not happen to their older colleagues. The Stanford team found the effect concentrated in exactly the occupations you would predict if AI were the driver — software development, customer service, and similar roles where a capable model can do a meaningful slice of what a junior person used to do.
Crucially, the researchers drew a distinction that most of the panicked coverage flattened. The declines showed up where AI automates work — replacing the task outright — but were muted in jobs where AI mostly augments a human, making them faster rather than unnecessary. That single distinction is the most useful thing in the entire report, because it turns a vague dread (“AI takes jobs”) into an actionable question (“is my role being automated or augmented?”). Those are very different futures, and they are already diverging in the payroll data.
Now the objection — and it is a serious one
Here is where a responsible reading has to slow down, because the most obvious alternative explanation is also a very good one. Something else happened in late 2022: the Federal Reserve began the sharpest interest-rate tightening cycle in four decades. Higher rates hit exactly the kind of speculative, growth-funded companies — tech startups especially — that had spent the pandemic hiring juniors by the thousand. When the cost of capital exploded, those were the first roles cut. On that reading, AI is not holding the knife at all. It is just standing next to the body.
This is not a fringe position. A paper from the Economic Innovation Group argued the more plausible explanation is a classic macroeconomic shock, not technological displacement — and warned that ordinary workforce aging can create a statistical illusion of a targeted hit to the youngest cohort when the real cause is a broad, across-the-board hiring slowdown. Named skeptics piled on: as Fortune reported, Google economists attributed the decline to interest rates, others to tech-sector overhiring and remote-work distortions, and Apollo’s Torsten Slok questioned whether there is an AI jobs crisis at all rather than an ordinary low-hire, low-fire market.
Why the argument hasn’t been settled by one side winning
The reason this debate has legs is that the Stanford authors did not just wave the objections away — they went back and tested them, which is what makes the exchange worth following rather than just picking a team. They re-ran the analysis excluding the tech sector, computer jobs, and remote-workable roles, and reported similar results — suggesting the pattern is not purely a tech-overhiring or work-from-home story. They controlled for firm-level economic shocks and still found entry-level hiring in AI-exposed jobs declining relative to less-exposed jobs inside the same companies. And Brynjolfsson offered the cleanest single rebuttal to the interest-rate theory: the most rate-sensitive occupations, like construction, have some of the lowest AI exposure — so if rates were the whole story, the damage should show up in different places than it does.
And yet the skeptics are not refuted either. The honest state of play was captured well by Stanford’s own policy institute, SIEPR, which acknowledged that isolating AI’s impact is empirically challenging given the other macroeconomic shocks hitting at the same time, and that this remains an open, active area of research. International data muddies it further: some studies of UK and US firms find entry-level contraction tied to AI adoption, while a Danish study released in late 2025 found no hiring difference between firms that adopted AI and those that did not. When rigorous studies in different countries disagree, the responsible conclusion is not “AI is guilty” or “AI is innocent.” It is “the effect, if it exists, is real enough to measure in some places and small enough to vanish in others.”
What to actually do with an unsettled answer
It would be easier if one side had won. But the useful move when the causation is genuinely contested is to notice what both sides agree on, because that part is not contested at all. Everyone — Brynjolfsson and his critics alike — agrees the entry-level rung has gotten thinner since 2022. They disagree only on why. And for a 23-year-old deciding what to do, or a manager deciding whether to hire one, the “why” matters less than the “what now.”
The automation-versus-augmentation split is the most durable guidance to come out of this whole argument, precisely because it holds regardless of which macro story is true. Roles that mostly execute codified, teachable tasks — the classic first job whose value was “cheap hands that will learn” — are exposed whether the immediate cause is a model or a rate hike, because both push companies to do more with fewer juniors. Roles where a young worker’s judgment, relationships, or physical presence augment a process are far safer. The tell is not the industry. It is whether the job’s core is a task a model can now do end to end.
There is also a quieter worry underneath the numbers that neither side disputes, and it may be the most important part. If companies stop hiring and training juniors — for whatever reason — the pipeline that produces senior people does not refill itself. You cannot skip the bottom rung and still have a ladder. Whether AI is the cause or merely the excuse, an economy that quietly stops investing in its least experienced workers is borrowing against its own future supply of experienced ones. That is the story worth watching, and it does not depend on winning the argument about blame.
So resist both of the tidy narratives on offer. “AI is destroying a generation’s careers” is running ahead of what the data can currently prove. “There’s nothing to see here, it’s just interest rates” is running away from something the data keeps stubbornly showing. The accurate version is less dramatic and more useful: the entry-level job is genuinely getting harder to find, the reasons are still being fought over by serious people with good evidence, and the smartest response — as a worker or an employer — is to bet on the things that stay true no matter who wins.
Related reading: Everyone’s Adopting AI. Almost Nobody’s Profiting. That’s the Real 2026 Story
