In August, Stanford's Digital Economy Lab updated its study of what AI is doing to jobs, using ADP payroll records covering 3.5 to 5 million workers a month. Employment of 22- to 25-year-olds in the occupations most exposed to AI now stands 19% below where it would be if it had kept pace with young workers in less-exposed jobs. A year earlier, that gap was 15% (Stanford Digital Economy Lab).
Experienced workers in the same occupations show no comparable gap.
And this isn't a layoff story. The decline comes mainly from reduced hiring, not from more people leaving. Separation rates actually fell, and among young workers they fell at least as much in the most exposed jobs. The authors call that "the opposite of what displacement would imply."
That is why this is easy to miss. Attrition looks normal. There are no layoffs to announce. The headcount report is fine. What's missing is a group of people who were never hired, and no dashboard counts people who aren't there.
What the data says about developers
Software development is one of the occupations the study singles out. In the November 2025 version of the paper, the authors reported that by September 2025, employment of software developers aged 22 to 25 had fallen nearly 20% from its peak in late 2022, while employment for older developers kept growing (Stanford, Nov 2025). The August update doesn't give a newer developer-only figure in its main text, so treat that number as a year old.
A separate study points the same way with different data. Harvard researchers Seyed Hosseini and Guy Lichtinger used LinkedIn résumé and job-posting data covering about 281,000 firms. At companies that adopted generative AI, junior employment fell roughly 9% relative to non-adopters six quarters after the start of 2023, while senior employment showed no break in trend. Again, the main channel was slower hiring, not more exits (Hosseini & Lichtinger).
There are honest caveats. Stanford says the effect is more pronounced in ADP's payroll data than in national surveys, which show the same direction with a smaller, noisier gap. The authors describe their results as "early, descriptive indicators" rather than proof of cause. An earlier note from the lab said the steep decline for young developers has "various factors" behind it, and that AI is not always the sole one (Stanford, Feb 2026). And an analysis on the Economic Innovation Group's Agglomerations blog, working from 2023 data, argued that the market for young computer-science graduates was holding up better than viral charts suggested (EIG Agglomerations).
The size of the effect is still being argued over. The direction is consistent: fewer juniors are being hired into AI-exposed work, and experienced people are not being pushed out.
Why the bill arrives in 2031
Here's the part the hiring numbers don't show.
One possible mechanism the Stanford authors propose is that AI substitutes for codified knowledge, the kind that can be written down, while complementing tacit knowledge, the kind "acquired through practice, mentorship, and repeated exposure to real situations." Occupations that rely on tacit knowledge are seeing faster employment growth for experienced workers.
Think about what that means for a software team. The skill AI is making more valuable is the one senior engineers carry: knowing why the billing service handles refunds the odd way it does, which migration will lock the table, what the business actually meant by the requirement. That knowledge is built over years of junior work. Small fixes, code reviews, on-call shifts, being wrong in front of someone who explains why. AI is absorbing a lot of that entry-level work. Cut junior hiring and you cut the supply line for the very thing that's getting more valuable.
Meanwhile, demand for software developers overall isn't going away. The U.S. Bureau of Labor Statistics projects employment of software developers, QA analysts and testers to grow 10% between 2025 and 2035, with about 106,100 openings a year (BLS). If every company makes the locally sensible choice to skip juniors now and hire experienced people later, they will all be bidding for the same pool later.
The arithmetic is simple. Someone hired at 23 in 2026 is a five-year engineer in 2031, roughly the experience we'd look for in someone owning a service or leading a small team. A hiring gap now becomes a mid-level and senior gap then. No study has measured that yet, because it hasn't happened. This is our read of where the data points.
AWS CEO Matt Garman was blunter last year. Replacing junior employees with AI, he said, is "the dumbest thing I've ever heard" (TechRadar).
Picture a mid-market company with a 30-person engineering team. This is a composite, not a specific client. In 2026 it pauses graduate hiring because AI tools let its seniors cover the routine tickets. Throughput holds. The budget looks better. By 2029, two senior engineers have left, as people do. The team goes to market for replacements and finds every other company that made the same call is looking too. The new hires are expensive and take months to learn the systems. The knowledge of why things work the way they do now lives in fewer heads, and the people who would have been ready to step up were never hired.
Nothing in that story was a bad decision at the time. That's the problem.
What to do about it, starting Monday
Hiring juniors as a favor won't fix this. What helps is redesigning entry-level work for a team that uses AI, so juniors still build judgment.
Keep a junior pipeline in the plan, and budget it as one. Treat it like any capability you'll need in five years, not as the first line to cut.
Make review the classroom. Pair juniors with seniors on structured code review, of their own AI-assisted work and of other people's. Review is where tacit knowledge changes hands.
Set norms for how juniors use AI. Ask it to explain, not only to produce. In a small controlled study we covered in an earlier post, engineers who used AI to ask conceptual questions scored far better on a follow-up quiz than those who handed it all the code writing.
Give juniors ownership of bounded pieces. A small service, a migration step, an on-call rotation with a senior behind them. Tacit knowledge comes from being responsible for real things.
Ask your delivery partners the same question. A partner staffed only with seniors and AI tools has the same 2031 problem you do, and you inherit it at handover.
This is the "right people" part of what we wrote about in our first post: tools amplify the team they're dropped into. In our own work, we pair AI tools with structured review and mentoring, so less experienced engineers learn the system while they ship, rather than only learning how to prompt.
If you're weighing a hiring freeze against your AI tool budget, let's talk.