Unscarcity
Sign in for free: Preamble (PDF, ebook & audiobook) + Forum access + Direct purchases Sign In

Unscarcity Research

The Vanishing Apprenticeship: Who Trains the Next Senior?

Goldman found AI's hiring drag lands hardest on juniors. But entry-level work wasn't cheap output. It was the factory that manufactured senior judgment.

15 min read 3397 words Updated August 2026 /a/vanishing-apprenticeship-rung

Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.

The Vanishing Apprenticeship: Who Trains the Next Senior?

Every firm that ever hired a junior was running two businesses at once, and it only ever charged for one of them.

The first business was output: the deck, the ticket, the test suite, the file review, the call that got answered. Cheap work by cheap people, priced accordingly. The second business was manufacturing. Somewhere in the middle of all that tedium, a person was being built. Five years of doing unglamorous things under real consequences, and out the other end came somebody with judgment — the ability to look at a situation and know, without being able to fully explain why, that something is off.

Nobody ever sent an invoice for the second business. It came free with the first. Which means that when a firm automates the first, it cancels the second by accident, and it does not find out for about a decade.

The receipts arrived in August

On August 19, 2026, Goldman Sachs published a Global Economics Comment asking whether AI is showing up in global labor data yet. The bank’s economists worked through employment growth across more than 800 occupations in a dozen developed economies. The headline conclusion was deliberately unspectacular: AI-related hiring headwinds are, in the report’s own framing, “clearly visible in official and unofficial employment data, but impacts are limited to a narrow set of industries and workers.”

Then you read which workers.

For the workforce as a whole, every 10-percentage-point increase in an occupation’s AI exposure was associated with roughly a 0.1-point drag on annual headcount growth in the US, Canada and France. For US entry-level roles, the same 10 points cost more than 0.2. In Australia, entry-level took a hit of over 0.6. Goldman’s own characterization of the gap, as reported by IBTimes, is that the effect on entry-level workers ran “more than three times larger than the effect across the broader workforce.”

The industry-level numbers are blunter still. US call-centre employment is running 39% below its historical trend, 33% below in Canada, 27% below in Germany, with software publishing, management consulting and advertising services all dragging under trend as well. And all of this is happening at AI adoption rates of only 15% to 20% across developed economies. This is what a fifth of the way in looks like.

Goldman is not alone, which is the part that should worry you. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the Stanford Digital Economy Lab ran ADP payroll microdata and found a 13% relative decline in employment for workers aged 22 to 25 in the most AI-exposed occupations, concentrated specifically where AI automates rather than augments. And SignalFire’s 2026 talent data shows the mechanism at the firm level: recent graduates are now 7% of Big Tech hires, down from 15% pre-pandemic, entry-level hiring at the twelve largest US tech companies is down roughly 65% against 2019, and startups have cut new-grad hiring from 30% to under 6%.

Three different datasets, three different methods, one shape. And SignalFire says the quiet part out loud: the squeeze comes from AI absorbing the boilerplate code, the unit tests and the routine debugging that juniors used to own.

Read that list again. Nobody is mourning boilerplate. But boilerplate was the syllabus.

Grunt work was a by-product factory

Here is the conceptual move that most of the coverage misses. The tasks being automated at the bottom of the ladder were never valuable primarily as output. They were valuable as reps.

The Stanford paper puts a useful frame on it. Young workers trade mostly in codified knowledge — the stuff you can write down, test on, and learn in a classroom, which is precisely the stuff a language model trained on the written record is best at. Experienced workers trade in tacit knowledge: intuition, situational read, the accumulated pattern-matching that nobody can articulate well enough to put in a manual. As Brynjolfsson put it, “what younger workers know overlaps with what LLMs can replace.”

That sounds like good news for anyone over thirty-five. It is not, and the reason is a single ugly dependency: tacit knowledge is a residue of doing codified work under consequence. It is not a separate skill you can be taught. It precipitates out of years of doing the boring thing and occasionally getting it wrong in front of someone who noticed. The moat that currently protects senior workers was dug by the very activity we are now automating.

So the ladder is not being shortened from the bottom. It is being disassembled from the bottom, and the rungs above it are made of the same material.

This is the same dynamic that the Substitution Threshold describes for individual tasks, running one level up. The threshold tells you when a machine becomes the cheapest reliable provider of a task. It does not tell you what else that task was quietly doing for you. Most tasks are load-bearing for something other than their output, and the accounting system has never had a column for it.

The lag is the entire problem

An honest objection: firms are not stupid. If gutting the junior tier creates a senior shortage, firms will notice and fix it.

They will notice. They will notice in about ten years, which is roughly the interval between hiring a graduate and having someone you can put in front of a client without supervision. That is longer than the average CEO tenure, vastly longer than any planning cycle a CFO models, and infinitely longer than the quarter in which the headcount decision gets made. The manager who cancels the graduate program collects the margin now. The bill lands on their successor’s successor, unitemized.

Worse, training has always been a positive externality that individual firms cannot capture. You spend five years and a lot of senior attention turning a graduate into someone useful, and then a competitor hires them with a 30% raise. Under wage employment, the rational move for any single firm has always been to let someone else train people and then poach them. What AI did was remove the last reason not to defect: the junior used to at least pay for themselves in output while learning. Now they don’t. The subsidy is naked, and nobody funds a naked subsidy.

Which means every firm defects, and the industry as a whole discovers, sometime in the mid-2030s, that the pool it was planning to poach from does not exist. You cannot hire your way out of a shortage of people who were never made. This is where the AI talent paradox — capital and compute abundant, elite human judgment the binding constraint — stops being a story about a few thousand researchers and becomes a story about the middle of every profession.

And note what this does to the labor data. The damage takes the form of jobs that were never posted, which is exactly the category official statistics are worst at seeing. The employment numbers count people who lost something they had. They have no field for the career that did not start, which is why the Labor Cliff keeps looking milder in the aggregate than it feels to anyone under twenty-six.

We have run this experiment before, twice

The strongest reason to take the delayed-deficit argument seriously is that it is not a prediction. It is a replication.

Surgery. In 2003, US graduate medical education imposed duty-hour restrictions on residents — a humane reform, badly needed, and one that reduced the hours trainees spent in operating rooms. Nothing broke. Patients did fine. Then a decade later, Mattar and colleagues surveyed fellowship program directors for Annals of Surgery and asked them what was walking through the door. The answer: 30% of incoming fellows could not independently perform a laparoscopic cholecystectomy, 38% showed no ownership of their patients, and 66% could not be left to operate unsupervised for thirty minutes of a major procedure. These were board-eligible surgeons after five years of training.

Ten years of invisible, then a deficit that arrives fully formed and cannot be patched with a workshop.

Aviation. The FAA’s Flight Deck Automation working group spent years on the question and published Operational Use of Flight Path Management Systems in September 2013: 279 pages, 29 findings. Automation made flying dramatically safer. It also produced documented degradation of manual flying skills, over-reliance on the automation, and pilots who had lost the habit of monitoring it. Manual handling or flight-control error showed up in about 60% of the accident, incident and observational data reviewed, and pilot overconfidence in automation contributed to roughly a quarter of the accidents examined.

Both statements are true simultaneously. The automation is better than the human. The human gets worse because of the automation. Any argument that treats those two facts as contradictory is not describing the world we live in.

The complication: the seniors are depreciating too

Here is where the tidy version of this argument — cut the juniors, get a shortage in 2036 — is too optimistic.

In August 2025, The Lancet Gastroenterology & Hepatology published what the researchers called the first documentation of a clinical AI “deskilling” effect. Nineteen experienced endoscopists, each with more than 2,000 colonoscopies behind them, were studied across 1,443 procedures performed without AI assistance — 795 before their centres adopted AI detection, 648 after. Their adenoma detection rate on unassisted colonoscopy fell from 28.4% to 22.4%.

Six percentage points of cancer-precursor detection, gone, in people whose expertise was already banked. Nobody took anything away from them. They simply stopped being the one looking.

So the stock is decaying at the same time the flow is being cut off. The pipeline argument understates the problem by assuming today’s seniors hold their value while we fail to make new ones. They don’t. Expertise is not a certificate; it is a perishable good maintained by use, and we have just built a technology whose entire value proposition is that you no longer have to use it.

“But won’t the AI just be the senior?”

This is the serious objection, and it deserves better than a wave.

Maybe judgment is exactly what the next model generation supplies, and the whole apprenticeship worry is a buggy-whip complaint. Perhaps in 2036 we won’t need a partner with twenty years of intuition, because the system will have more pattern coverage than any human ever accumulated.

Two problems.

The first is verification. Someone has to be able to tell when the machine is wrong, and that capacity is precisely the thing being eroded at both ends of the pipeline. Where output can be checked mechanically — a proof that compiles, a test that passes — this is fine, and the verifiability premium explains why those domains flipped first. But most professional work is not mechanically checkable. It is checked by a person who has seen enough to be suspicious, and suspicion is the last thing you learn.

The second is accountability. In licensed work, “reliable” has never only meant accurate; it has meant that a specific human absorbs the loss when the answer is wrong. That is the liability gap, and it is why a model can already out-read your radiologist without replacing them. The person signing has to be competent to sign, or the signature is theater.

But here is the part that should end the argument. If the objection is right, it doesn’t rescue anybody. If AI genuinely supplies senior judgment by 2036, then the senior tier goes too — just later, and with worse severance. The optimistic case and the pessimistic case converge on the same labor market. They differ only on which decade the middle of the profession stops existing. Any plan that depends on which one is correct is not a plan.

Experience becomes the scarce input

The book’s argument is that the Labor Cliff arrives not as a single event but as the serial removal of the economic reason to employ a human. The apprenticeship story is the sharpest version of that, because it is the case where automating a task destroys the very capability that made the task worth supervising.

Once output is abundant, what stays scarce is the thing you can only make by putting a person in a situation with real stakes for a long time. Experience is the last input that has no supply curve. You cannot mint it, import it, or fine-tune it. And the way our economy is structured, nobody has any private incentive to produce it, because the producer never captures the return.

That is not a market failure to be patched with a tax credit for internships. It is the market working correctly against the wrong objective function. A wage economy pays for output. Training is a cost that yields output later, to somebody else. Of course it gets cut. It was always going to get cut the moment the output subsidy disappeared, and AI is simply what made that moment arrive on a specific Tuesday in 2026.

The Unscarcity read

This is where the framework earns its keep, because it addresses the exact failure the wage system produces.

The Foundation removes the requirement that learning be profitable in the moment. If food, shelter, healthcare, energy and compute are unconditional for every Resident, then a twenty-three-year-old spending four years becoming competent does not need to be net-productive on day one to be permitted to exist. The apprenticeship stops being a subsidy a firm has to justify to a board and becomes something a person is simply allowed to do. That single change kills the free-rider problem, because there is no longer a rival firm capturing a return the trainer paid for.

Impact prices the thing wages structurally cannot. Under an employment contract, the senior who spends an afternoon teaching instead of shipping has destroyed value on that day’s ledger. Under Impact, transferring capability to another person is a contribution to the Frontier and is measured as one. And because Impact decays, a master cannot hoard reputational capital and coast on it; the only way to hold standing is to keep contributing, which in a knowledge domain overwhelmingly means keeping other people’s competence current. The system pays for mentorship on purpose rather than tolerating it as overhead.

The Guilds put the ladder back inside an institution that survives its members. The book’s Mission and Frontier Guilds run explicitly on Apprentice → Contributor → Steward → Mission Guardian, and the point of that structure is that the entity holding the training obligation is the one that persists long enough to need the outcome. A firm optimizing a quarter cannot rationally invest in 2036. A guild whose entire reason to exist is domain capability has no other move. And Civic Service does at the societal level what the corporate graduate program used to do accidentally: it puts every young adult into real work with real consequences, deliberately, because the civilization needs the people, not because a P&L cleared.

None of this is nostalgia for grunt work. Boilerplate was miserable and the machines are welcome to it. The argument is narrower and harder: the misery was doing a second job we never named, and we have decommissioned the factory without checking what else it was making.

The bill comes addressed to nobody

The best thing about the entry-level squeeze, from a CFO’s perspective, is that it generates no press cycle. No WARN notice, no severance line, no crying-in-the-parking-lot photo. You simply do not open the requisition. Goldman’s 0.2 points of annual drag is not a wave of firings; it is a hiring page that quietly stopped refreshing, multiplied across 800 occupations and several continents.

Which is why almost nobody is treating it as an emergency. The costs are dispersed across a cohort that has no bargaining power, and the consequences are deferred past the horizon of everyone empowered to act. It is the most perfectly structured collective-action failure the labor market has produced in fifty years, and it is running right now, in public, with the data published.

The class of 2026 is not just being denied jobs. It is being denied the ten years of consequence that would have turned it into the people we will need in 2036 — and those people are the only ones who could tell us whether the machines running everything by then are getting it right.

That is the case Unscarcity makes: the floor is not charity for the displaced. It is the infrastructure that lets a civilization keep producing competent adults after the economy stops paying for the privilege. Build it while there are still people around who remember how the job is actually done.

Sources

Share this article:

Share Share Share