Cutting junior hiring is the correct decision for almost every company that made it.
The popular read blames short-sighted CFOs. That read is wrong, and it is comfortable, so it circulates. Every one of those decisions clears its own business case. The damage lands somewhere the business case never looks. Seniors are not found. Seniors are trained, and the training was always paid for in grunt work.
Every engineering leader I talk to says the same two sentences, usually in one breath. "We only hire seniors now." And: "it's so hard to find seniors." Nobody laughs. The joke is structural, so it does not register as one.
The data stopped being anecdotal
Stanford economists ran ADP payroll microdata, real employment records for millions of workers, and found the canaries.
Two details in the Stanford paper matter more than the headline. The adjustment came through employment, not wages. That is the signature of a junior nobody hires, rather than a junior somebody reprices. And the drop concentrated where AI automates instead of augments. Where the models assist, young employment held.
The hiring side cuts the same shape. Mid and senior hiring rebounded through the exact period new-grad hiring kept falling. The market recovered. The entry level did not. The whole ladder is shorter too, and "software engineer" titles are down 49% from before the pandemic.
Read the last ledger row again. The major sold to a generation as the safe bet now runs above an average that includes art history.
The intern paid tuition in grunt work
The apprenticeship deal was never charity. A junior was cheap labor at the bottom of the task ladder, and the grunt work was the tuition.
The arrangement is old and it is studied. Epstein's work on medieval guilds shows apprenticeship solved a market failure in training investment. Masters trained because enforceable multi-year contracts let them recoup the cost from the apprentice's later productive years. The apprentice paid with underpriced labor. The master paid with knowledge. The contract stopped either side from defecting early.
England thought this important enough to legislate. The Statute of Artificers made a completed seven-year apprenticeship a legal requirement to practice most trades, from 1563 until 1814. There is a respectable case (de la Croix, Doepke and Mokyr) that the institution explains a real chunk of why pre-industrial Europe pulled ahead. Dense person-to-person transmission of tacit skill, between people who were not relatives.
Skills transmission is infrastructure, not a perk. It has always needed scaffolding, because it has never paid for itself on a quarterly ledger.
Then the machine started doing the grunt work, and the tuition currency collapsed. An LLM produces intern-quality output on intern-shaped tasks with no onboarding, no equity, and no manager's Tuesday. The seats stopped making quarterly sense. The seats went away.
The economics has a name
This outcome is individually rational, and that is what makes it collectively stupid.
Economists have known since Pigou in 1912, formalized by Becker that a frictionless labor market kills general training. Train someone in portable skills and a competitor who trained nobody poaches them. You cannot match the offer, because you are still amortizing the training. Call it the poaching externality.
So why did tech firms train juniors anyway, for seventy years? Acemoglu and Pischke's answer is frictions. Switching jobs was costly. Information about outside offers was bad. Wages were compressed. General skills behaved like firm-specific ones, and the investment paid back quietly.
Now run the 2025 update, term by term. Remote work made switching nearly frictionless. Salary transparency and levels.fyi killed the information asymmetry. AI zeroed out the junior's first-year output, the thing that used to offset training cost while the judgment grew.
Every term pushes toward the Becker corner. Training is nobody's rational move and everybody's necessity. That is a market failure of the textbook kind, and exhortation does not fix those.
The objections, taken seriously
Two pushbacks deserve airtime.
The first: AI is the best tutor a junior ever had, so the pipeline accelerates instead of shrinking. A motivated 23-year-old with an agent learns in months what took years. Partly true, and it misses what seniority is made of. The tutor teaches skills. Judgment grows by owning consequences. You ship the migration that pages you at 3am. You defend the estimate that slipped. You unwind your own bad abstraction a year later. AI compresses skill acquisition and does nothing for consequence acquisition. Companies just removed the seats where consequences came in junior sizes.
The second: this is just the hiring cycle, 2022 was a bubble and everything reverted. The SignalFire asymmetry answers it. Senior hiring recovered. New-grad hiring kept falling through the recovery. A cycle sinks all boats. This one dropped a single deck.
The lead time keeps the bill invisible until it is enormous. Seniority takes five to ten years to grow. The cohort nobody hired in 2023-2025 is the missing staff engineers of 2032 and the missing principals of 2035. They arrive exactly as the seniors everyone hires today start aging out. Demography is the slowest, surest form of technical debt.
What restocking looks like
The last time skills transmission faced this incentive structure, the answer was institutions rather than virtue. The existing ones are instructive.
- Medicine. Residency is a funded, mandatory apprenticeship. It survives because no hospital can skip it and poach instead.
- Germany. The dual system splits the bill between firms and the state, and youth unemployment stays structurally low.
- The UK. The 2017 apprenticeship levy taxes payroll and refunds it to whoever trains. Clumsy, gamed, and directionally correct.
The software version writes itself in some combination. Agent-era residencies, where juniors own small blast-radius systems with real consequences and AI does the typing. Training booked as infrastructure spend, with retention contracts doing what guild indentures did. And if the free-riding persists, someone will reach for the levy. The market failure is textbook. So are the remedies.
If you are the junior in this story, the advice compresses to one line. Manufacture your own consequences. Own something end to end anywhere that will let you. An OSS package with real users. A side project with real invoices. The unglamorous internal tool nobody else wants. Ownership is the input the market stopped providing and still pays for.
The companies that harvest seniors trained on someone else's payroll run the oldest free-rider play in economics. They started it at the exact moment the pool stopped refilling. Ask them, in 2035, who trained their seniors. Someone did. It will not have been them.