Stanford researchers tracked ADP payroll microdata for every high-AI-exposure occupation since ChatGPT launched. Workers aged twenty-two to twenty-five lost nearly twenty percent of their jobs. Workers over thirty in the same occupations grew six to twelve percent. AI displacement is not an industry question. It is an age question.
A week ago, this journal asked a question: where do the experienced workers of 2036 come from, if the apprenticeship pipeline of 2026 is being dismantled? Stanford just answered it with data. The answer is: they might not.
Researchers at the Stanford Institute for Economic Policy Research — Erik Brynjolfsson, Alex Chandar, and Rui Chen — published a study titled "Canaries in the Coal Mine" using ADP payroll microdata covering millions of workers. They tracked employment changes in every occupation with high AI exposure since ChatGPT launched in late 2022. The occupations included software development, customer service, data analysis, and technical support — the fields where AI tools have been deployed fastest and most aggressively.
The headline finding is not about industries or occupations. It is about age.
The Numbers
Workers aged twenty-two to twenty-five experienced a relative employment decline of thirteen to sixteen percent across all high-AI-exposure occupations. In software development specifically, the decline for that age group approached twenty percent. These are not unemployment statistics — they are employment shares. The young workers' slice of the workforce in these fields contracted by a fifth in roughly three years.
Workers over thirty in the same occupations — the same companies, the same job families, the same industry classifications — saw employment grow six to twelve percent over the same period.
The same occupations. Opposite directions. Sorted entirely by age.
The study was presented at the 2026 SIEPR Summit in March. Its methodology relies on the ADP payroll dataset, which covers roughly one-fifth of the U.S. private workforce and tracks individuals at the establishment level — making it possible to distinguish between companies shrinking overall and companies replacing one age cohort with another.
Not What It Looks Like in Aggregate
The aggregate statistics hide this. The Bureau of Labor Statistics reported a ninety-two-thousand-job decline in February — the headline that dominated financial news. But aggregate payroll data does not break down by age within occupation. It shows net change. A company that fires five twenty-three-year-old developers and promotes three thirty-five-year-olds into expanded roles shows a net loss of two. The number is accurate. It obscures the mechanism.
The Dallas Federal Reserve identified the mechanism earlier this year: AI simultaneously aids and replaces workers within the same industry. Computer systems design employment fell five percent since ChatGPT launched, but wages in the same field rose 16.7 percent — more than double the national average of 7.5 percent. Fewer workers, higher pay. The Stanford data reveals what the Dallas Fed data implied: the workers being removed and the workers being paid more are different people, sorted by how long they have been doing the work.
This is the experience divide operating at the individual level. Entry-level workers perform codifiable tasks — the work that can be described in a manual, followed from a runbook, evaluated against explicit criteria. AI performs these tasks at scale. Experienced workers perform judgment tasks — the work that requires pattern recognition built through years of handling exceptions, making mistakes, and developing intuitions that resist formal description. AI augments these tasks. The same technology, applied to the same occupation, displaces one group and empowers the other.
The dividing line is not what you do. It is how long you have been doing it.
The Tax
An inexperience tax is a cost that accrues silently, compounds over time, and arrives as a crisis.
Every senior engineer was once a junior engineer who did not know what they were doing. Every experienced nurse practitioner was once a new graduate who needed supervision on routine procedures. Every seasoned trader was once an analyst who misread a chart. The tacit knowledge that makes experienced workers irreplaceable — the knowledge that AI augments rather than replaces — is not taught in school. It is built through years of performing the codifiable tasks that AI now handles.
The apprenticeship tasks are not the point of the job. They are the mechanism through which workers become the point. Data entry teaches what clean data looks like. Level-one support teaches what users actually struggle with. Writing test cases teaches where software breaks. Reviewing contracts teaches which clauses matter. These are the tasks that AI performs faster, cheaper, and more consistently than any twenty-three-year-old. They are also the tasks that produce the thirty-five-year-old whom AI cannot replace.
Cut the entry-level pipeline and the cost does not appear immediately. The experienced workers are still there. They are, in fact, more productive than ever — augmented by the same tools that displaced their younger colleagues. The wage data confirms this: compensation is rising for the experienced cohort. The organization looks healthier in the short term. Leaner. More efficient. The stock price reflects it.
The tax comes due in five to ten years, when the current experienced cohort retires, burns out, or moves on — and the pipeline that would have produced their replacements was cut a decade earlier. The twenty-two-year-old who would have spent 2024 through 2028 learning the craft by doing the boring parts was never hired. The thirty-two-year-old who would have carried institutional knowledge through the next decade does not exist. The organization has a gap it cannot fill by hiring, because the candidates with the requisite experience were never given the chance to acquire it.
The Feedback Loop
The economics create a self-reinforcing cycle. AI displaces junior workers, making experienced workers scarcer. Scarce experienced workers command higher wages, increasing labor costs. Higher labor costs accelerate AI adoption for any remaining codifiable tasks, further reducing the entry-level pipeline. Each turn of the cycle makes experienced workers more valuable and the mechanism for producing them narrower.
The World Economic Forum projects that 170 million new roles will be created globally while 92 million are displaced — a net gain of 78 million jobs. The number is reassuring until you examine the composition. The new roles require the judgment, creativity, and cross-domain thinking that come from years of experience. The displaced roles are precisely the entry-level positions that build that experience. The net number is positive. The pipeline that connects one to the other is shrinking.
Mercer's 2026 Global Talent Trends survey found that employee fear of AI job loss jumped from 28 percent to 40 percent in a single year. Among workers aged twenty-two to twenty-five, the fear tracks reality — the SIEPR data confirms they are being displaced at roughly the rate they fear. Among workers over thirty, the fear exceeds reality — they are gaining employment share, not losing it. The psychological tax is distributed differently from the economic one, but both concentrate on the youngest workers.
What the Number Actually Measures
The twenty percent figure for young software developers is a relative employment decline, not an unemployment rate. Some of those workers found positions in other fields. Some went to graduate school. Some left the workforce. The SIEPR study tracks employment shares within occupational categories — it measures whether the composition of the workforce shifted, not whether individuals found other work.
But the number measures something that individual outcomes do not capture: the field's capacity to reproduce itself. A profession that stops absorbing new entrants does not collapse immediately. It ages. The median experience level rises. The knowledge base narrows to what the current cohort knows. The diversity of approaches — the fresh perspectives, the unconventional methods, the naive questions that occasionally reveal something the experts missed — contracts. The organization becomes simultaneously more expert and more fragile.
Computer systems design lost five percent of its total employment while wages rose seventeen percent. The Stanford data reveals that the loss was concentrated in workers under twenty-five and the gains were concentrated in workers over thirty. This is not a shrinking industry. It is an industry that has stopped replenishing itself.
The word for a system that consumes its seed corn is not efficient. It is extractive. The inexperience tax is what an industry pays when it optimizes for the present at the cost of the future — and the invoice arrives only after it is too late to change the decision.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)