The Missing First Rung
Stanford tracks the payroll of five million Americans a month. Employment for 22-to-25-year-olds in AI-exposed jobs sits nineteen percent below where it should be. Nobody is being laid off — they are simply not being hired. Those are entirely different problems.
Everyone has an answer to whether AI is destroying jobs, and nobody actually knows. The problem is that the answers almost always come from one of two bad sources: executive statements, or counts of job postings. The first tells you what a company wants to say. The second tells you what a company advertised. Neither tells you what a company actually did.
That is why the Stanford Digital Economy Lab report belongs in a different category. They work with ADP, the largest payroll processor in the United States. What they hold is neither opinion nor advertisement but real payment data: the wages of roughly five million workers a month — who got paid, how much, and when it stopped.
The paper is called "Canaries in the Coal Mine?" — after the bird miners carried underground. If there was gas, the bird noticed first. The title is itself a claim: this is not what is happening to everyone yet, this is the group that notices first.
The latest version came out on August 12 of this year. These are the numbers I took from it.
One number
Employment for workers aged 22 to 25 in AI-exposed occupations is nineteen percent below where it should be.
"Where it should be" is the important part of that sentence, because this is not a simple decline. The researchers compute it this way: how did employment change for people of the same age in occupations that are not AI-exposed? If the exposed group had moved the same way, the gap would be zero. The gap came out at nineteen percent.
So this is not an economic slowdown. A slowdown hits every occupation roughly evenly. This is a specific kind of work disappearing for a specific age group.
The second number says even more than the first: experienced workers show no comparable gap at all. Same occupations, same companies, same AI — and employment for people over thirty-five has not fallen. In places it grew.
One technology is having opposite effects on the two ends of the same career. Why is the real question.
Not layoffs — non-hiring
The most important finding in the report is not the size of the number. It is the mechanism.
The decline is not happening through layoffs. Young people already employed are not being let go. It is happening through reduced hiring — requisitions never opened, internship programs quietly cancelled, the unannounced decision that there will be no junior class this year.
That distinction is not academic. It produces a completely different problem.
A layoff is visible. It makes the news, it lands in the statistics, a union responds, a politician comments. Non-hiring is not an event at all. Nobody announces "today we did not hire three people." A line simply fails to appear in a budget. Which is exactly why this process can run for years without attracting attention — as it has already been running.
There is a second consequence. A laid-off worker looks for a job and usually finds one, because they have experience. A person who was never hired cannot begin accumulating experience. If the first rung of the ladder has been removed, the question becomes how anyone reaches the second — and right now nobody has a good answer.
Substitution versus complementation
Now to the part of the report that gets the least attention, because in practical terms it is the most valuable.
The decline is not everywhere AI is used. The decline is concentrated where AI substitutes for human tasks. Where AI complements the worker, employment is flat or rising — especially for experienced workers.
This is worth pausing on, because it splits a single job title in two.
Say your work is writing boilerplate against a clear specification. A ticket arrives: add validation to this form, add a filter to this endpoint. The requirement is explicit, the solution is standard, the result is easy to verify. That is the substitution column. AI does this well and does it cheaply.
Now a different kind of work. You get a vague complaint: payments are sometimes going through twice. That is not a ticket, that is a symptom. Someone has to discover that a provider is sending the callback twice, trace the retry logic to find where the idempotency key got dropped, and confirm the fix will not break an old mobile client. AI is genuinely useful here — but only after someone has framed the question correctly. That is the complementation column.
A year ago these were one job. A junior came in, started on the first kind of work, and grew into the second over two or three years. Boilerplate was the training ground.
The training ground has now been automated. The job is splitting in two.
Why this is not only about programmers
The report does not single out software engineers. Its list of AI-exposed occupations is much broader: customer support, bookkeeping, marketing copy, translation, data entry, legal document preparation.
The pattern is the same across all of them. Within any profession, the entry-level work tends to be the most codified, the most rule-bound, the easiest to explain — which is precisely why it gets handed to the new person. And precisely why AI takes it first.
So this is not a story about programmers being finished. It is a story about the entrance to a profession narrowing, and it is happening in dozens of professions at once.
What to hold carefully
Three things, because these numbers are being flattened very fast online.
First: the report itself does not find mass unemployment. That is literally its first of six facts — no evidence of widespread, economy-wide job displacement. This matters, because the same study is being cited to support headlines claiming AI took everyone's job. It does not say that. It found a narrow, specific effect in one age band, in one kind of work.
Second: this data is American. ADP is a US provider. The situation in Uzbekistan, or in developing markets generally, may differ — wage structures, outsourcing models, and local demand are all different. The pattern will probably travel, but "probably" is a guess, not a measurement. Transplanting a number like this directly into a local context is a mistake.
Third: correlation is not cause. This period overlaps with large tech-sector layoffs, high interest rates, and the correction of post-pandemic overhiring. The researchers try to control for these and isolate the AI effect — but this is econometrics, not a laboratory experiment. AI is plausibly one cause of the decline rather than the cause.
The conclusion I hold firmly is narrower: there is a gap in real payroll data, steadily widening for over a year, confined to one age group, and concentrated exactly where AI substitutes rather than assists. That is not proof. It is a signal — and a considerably more valuable signal than a press release.
What to actually do
The question that remains is what changes once you know this, because a diagnosis is not useful on its own.
The main change is in priority. If most of your work sits in the first column — explicit requirements, standard solutions — then spending your time learning to write code faster is no longer the best investment available. AI improves along that axis faster than you do.
The value moved to the second column, and the skills there are different in kind: turning a vague complaint into a precise problem, seeing in advance where a system breaks under load, looking at AI-generated code and saying "there is no transaction here," and knowing what not to build.
And there is one uncomfortable truth. The skills in the second column were traditionally learned by working in the first. Since that path has narrowed, they now have to be sought deliberately: take real ownership on a real project, fix a broken system, put your own side project into production and watch it fail. Nobody is going to hand you that as a ticket.
The first rung has been removed. That does not mean the ladder cannot be climbed — it means you have to build a rung yourself.
Source: Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence", Stanford Digital Economy Lab, August 12, 2026 version. Also: the Stanford Canaries Dashboard and IEEE Spectrum for NACE Job Outlook 2026 and SignalFire figures.
