# Every Junior Job Was Secretly Two Products > Source: > Published: 2026-08-02 12:00:00+00:00 # Every Junior Job Was Secretly Two Products AI now does the work juniors used to learn on, so firms have rationally stopped hiring juniors. The bill arrives around 2035, roughly a decade after the last moment anyone can act on it. ## Contents [The job was two products wearing one salary](#the-job-was-two-products-wearing-one-salary)[This stopped being a forecast about two years ago](#this-stopped-being-a-forecast-about-two-years-ago)[No one in this story is being stupid](#no-one-in-this-story-is-being-stupid)[A full forest does not miss its saplings](#a-full-forest-does-not-miss-its-saplings)["Come on. This was interest rates and return-to-office."](#come-on-this-was-interest-rates-and-return-to-office)[Medicine ran this experiment once and published the answer](#medicine-ran-this-experiment-once-and-published-the-answer)[The version where the problem eats itself](#the-version-where-the-problem-eats-itself)[What would prove this post wrong](#what-would-prove-this-post-wrong) I want to walk you through the worst year of a lawyer's life, because the industry has spent the last three years deleting it, and I do not think anyone has priced what it was for. It is 2019, and a first-year associate at a big firm is billing two thousand hours. Here is what is inside those hours. Document review, which means reading strangers' emails for relevance, ten hours at a stretch. First drafts of memos that a partner will rewrite so completely the draft functions as a suggestion. Citation checking. Proofreading contract language against a playbook. The associate does nothing all year that a client would recognize as legal brilliance, the firm bills the time out at several hundred dollars an hour anyway, and everyone involved understands the arrangement is a little absurd. Now it is 2026, and a model [does the document review](../03-domains/cognitive/law.html). It drafts the memo. The partner gets a better first pass than the first-year ever produced, in minutes, for approximately nothing, and the client stops paying several hundred dollars an hour for someone to learn on their dime. Everybody wins. Except the arrangement was only absurd if you believed the billing system, and the billing system was lying about what that year was. ## The job was two products wearing one salary[#](#the-job-was-two-products-wearing-one-salary) Every junior knowledge job was a bundle. Product one: the memo, the review pass, the model in the spreadsheet, the thing the firm wanted this week. Product two: the slow conversion of a graduate into someone whose judgment you can charge for, produced as a side effect of doing product one ten thousand times while a senior person checked. The firm only ever paid for product one. Product two rode along free, because there was no way to get the memo without accidentally training the person who wrote it. Look at what the buyer is looking at: the rolled-up memo poking out of one end. The graduate grinning out of the other end is not on the receipt and never has been. One parcel, one coin, two products. Call this **the Bundle**, because every judgment profession had its own version. Law bundled training into document review. Consulting bundled it into research memos and slide decks. Software bundled it into bug tickets and code review. Audit bundled it into ticking and tying. Nobody budgeted for training the next generation of seniors, anywhere, ever, because the Bundle meant nobody had to. AI splits the Bundle. The firm can now buy the memo alone. And here is the sentence [the corpus](../02-games/4-labor.html) builds this entire topic on: **the training was never the product, it was a positive externality of inefficient production, and efficiency destroyed it.** Nobody decided to stop training juniors. They decided to stop buying memos the expensive way, and the training was hiding inside the price. ## This stopped being a forecast about two years ago[#](#this-stopped-being-a-forecast-about-two-years-ago) Start with the strangest number in the US labor market. For as long as anyone has measured it, a fresh college degree bought you a lower unemployment rate than the workforce at large. That premium is gone. Recent graduates aged 22 to 27 were unemployed at 5.6% in March 2026, against about 4.3% for all workers, per the NY Fed's recent-graduates series (May 2026 release). The gap inverted around 2023 and is now the widest on record, and the underemployment rate for the same group is 41.5%. Then look at where the doors are. Indeed Hiring Lab (July 2026): US postings asking for five or more years of experience are up 14.7% year over year, while postings asking for zero to one year are down 7.5% and have been falling since 2022. In software, the most seniority-skewed field Indeed tracks, 69.3% of Q1 2026 postings were senior-level. The entry-level share of software postings was 4.5%. Out of every twenty software job listings in America, barely one is addressed to a person at the start of a career, in the occupation a decade of advice told every teenager to enter. At the big tech firms, new graduates are now about 7% of hires, down more than half from 2019 (SignalFire talent report, 2026). The cleanest research result matches. Brynjolfsson, Chandar, and Chen, working with ADP payroll microdata (the "Canaries in the Coal Mine" paper, revised November 2025), find employment for 22 to 25 year olds in the most AI-exposed occupations down about 16% relative to less-exposed peers, while older workers in the same occupations held steady. And the decline arrived through hiring, not layoffs. No headlines, no severance announcements. Firms just stopped opening the door. This is not an American quirk. UK graduate postings fell below 10,000 in January 2026, the first time since the series began in 2016 (Adzuna, via Bloomberg, February 2026), and KPMG UK cut its graduate intake 29% between 2023 and 2025 (Financial Times, 2025). And law, the profession that ran the purest version of the Bundle, shows the subtlest version of the break: AmLaw 200 firms hired 7,489 associates straight from law school in 2022 and 7,426 in 2025 (SurePoint hiring report, July 2026). Flat, across four years in which those firms grew revenue by tens of billions. A commons failure does not look like a massacre from the outside. It looks like a rung that quietly stops scaling with the business standing on it. ## No one in this story is being stupid[#](#no-one-in-this-story-is-being-stupid) It would be easier if there were a villain. There is only arithmetic. A firm that trains a junior pays the whole cost: years of salary against work a model now does better, plus the senior hours spent checking it. The benefit, a formed professional a decade later, is captured mostly by whoever employs that professional in year ten, which is usually somebody else. Training was always a terrible private investment. The Bundle forced firms to make it anyway, because the memo and the training could not be purchased separately. Unbundled, declining to train is the correct decision for every firm individually and a catastrophe for the profession collectively. That is the textbook definition of a [commons tragedy](../02-games/4-labor.html), this time in human capital, and the reason it deserves a blog post is its shape: the damage stays invisible for the entire window in which it is reversible. ## A full forest does not miss its saplings[#](#a-full-forest-does-not-miss-its-saplings) Picture the stock of senior professionals as an old-growth forest. Stand in it today and everything looks magnificent. Canopy closed, timber everywhere, no shortage of shade. Every partner, staff engineer, and audit director trained under the old Bundle is still working, and will be for twenty more years. The saplings are what got cut. And a forest with no saplings is indistinguishable from a healthy forest for about two decades. The big guy leaning on the trunk is every managing partner in 2029, and nothing he can see from there is wrong. The stumps the small one is pointing at are the entire crisis, and they are ankle-high. This is a stock-and-flow trap. The senior stock is enormous and depreciates over decades. The junior flow is what collapsed, and the flow is invisible in daily experience. For most of the next decade, every firm's lived reality will be senior abundance: experts available, projects staffed, nothing visibly broken. The people pointing at the missing saplings will appear to be wrong the entire time the problem is still fixable. Markets do carry a correction, and it is worth being precise about why it misfires here. When seniors finally get scarce, senior wages will spike, and the return to becoming one will rise. But producing a senior professional takes roughly a decade. The price signal arrives around 2035, which is when the cohort that should have been hired in 2026 through 2028 would have been finishing its formation. **Cohorts are not inventory. You cannot backfill 2028's missing associates in 2035.** So the corpus makes a checkable prediction: senior-to-junior wage ratios in exposed professions widen through the early 2030s, and firms respond not by reviving junior hiring but by poaching, thinning the senior layer, and redefining "senior" downward. ## "Come on. This was interest rates and return-to-office."[#](#come-on-this-was-interest-rates-and-return-to-office) This is the best objection, and it is partly right, so it gets the full treatment rather than a wave. Three confounds are real. First, the rate cycle: junior hiring is the most cyclically sensitive slice of white-collar employment, the 2022 to 2023 rate shock was the sharpest monetary tightening in four decades, and an EIG working paper (January 2026) found postings in AI-exposed fields falling for both juniors and seniors starting about six months before ChatGPT existed. Some of this is mean reversion from the 2021 hiring bubble. Second, remote work: a NY Fed analysis (June 2026) attributes about 64% of the post-Covid rise in young graduate unemployment to remote work, on the logic that firms will not hire novices into distributed teams where nobody can watch them and correct them. Third, composition: healthcare, government, and hospitality drove most 2024 and 2025 US job growth, and none of them hire many graduates. The Canaries authors answered the rate objection directly (February 2026 update): AI-exposed occupations are on average less rate-sensitive than unexposed ones, the most-exposed jobs declined more within both high and low rate-sensitivity groups, and they concede the pre-2024 weakness was macro. [The corpus's own split](../02-games/4-labor.html): roughly 50% substitution, 30% cycle, 20% work organization, held with low confidence and flagged as the number in the whole document most likely to be revised. But notice something about the remote-work story before you file it as a rebuttal. Its mechanism is that distributed teams broke the informal watch-and-correct loop that made juniors worth hiring. That is not an alternative to the apprenticeship problem. That is the apprenticeship problem with a different culprit. All three explanations describe firms declining to buy the same thing, training, and they differ only on why it stopped being worth buying. Only the AI version fails to reverse on its own. Which is convenient, because the discriminating experiment is already scheduled. White-collar hiring will recover on some cycle, probably around 2027 or 2028. If junior hiring recovers with it, the cycle was the story and this post overweighted AI. If aggregate hiring recovers and juniors do not, the substitution share was large. Call it **the Recovery Test**, and note that software is running an early preview with an ugly result: software postings finally rebounded in 2026, and 71% of the increase was senior roles (Indeed Hiring Lab, July 2026). The demand came back. The bottom rung did not come back with it. That drawing is what a two-tier recovery looks like. The two people walking in at the top are the senior hires, and the staircase under them is intact from the fourth step up. The graduate at the bottom is not facing a closed door. Just a gap that no one in the building considers their job. ## Medicine ran this experiment once and published the answer[#](#medicine-ran-this-experiment-once-and-published-the-answer) One profession hit this exact failure a century ago. Hospitals discovered that training doctors was expensive, that trained doctors walk out the door, and that no individual hospital would fund residencies at the scale the profession needed. The American solution was to stop pretending firms would pay for a public good: Medicare now funds graduate medical education at about $21 billion a year, supporting roughly 116,000 residents (CRS, FY2023 figures). Society decided the manufacture of seniors was infrastructure, and bought it directly. Notice who is tipping the jar in that drawing. Not a hospital. A crowd, all lifting together, pouring into a funnel no one of them owns. The queue of small doctors walking in the front door is what twenty-one billion dollars a year of collectively purchased training looks like. The template comes with a warning label. The Balanced Budget Act of 1997 froze the number of funded slots at 1996 levels, and the US has spent nearly three decades arguing about physician shortages since. Collective funding works. Collective funding also turns the width of a profession's pipeline into a budget line that someone can freeze. The residency model is what "someone buys the training as training" looks like when it is actually built. [The corpus lists](../06-uncertainties/apprenticeship-gap.html) the responses that could close the gap, and they run at different speeds: firms repricing their own pipelines is fastest, guild mandates like the medical model are slowest. So here is the scorecard as of August 2026. No professional body has mandated training ratios in response to AI. No firm that cut a graduate intake has publicly restored it. The most concrete response on record is Deloitte UK redesigning its three-year audit graduate programme, effective September 2026 (ICAEW, March 2026), around a first-year who directs and checks model output instead of producing it. One firm, repricing privately, exactly the fastest response type arriving first, and so far arriving alone. ## The version where the problem eats itself[#](#the-version-where-the-problem-eats-itself) There is one path where all of the above becomes a footnote, and it is worth stating in its strongest form because it is the only branch someone can deliberately build. The pessimistic case assumes expertise requires the decade of hours, and AI removed the hours. But the hours were never the point. The point was the feedback: ten thousand drafts, each corrected by someone who knew better. Education research has known since Bloom's two-sigma studies that individual tutoring produces about two standard deviations of improvement over classroom instruction, and the constraint was always that tutors do not scale. If AI can deliver dense, expert-quality feedback on real work at a volume no human mentor ever managed, [the novice-to-expert path gets shorter](../03-domains/cognitive/education.html), not broken. Expertise becomes cheaper to manufacture at exactly the moment it becomes scarce, and the shortage cancels itself. That stack of corrected pages next to the desk is taller than the robot, and it represents more feedback than any human mentor delivered across a whole apprenticeship. The picture is genuinely hopeful. Two problems with it. First, seniority is a bundle too: skill, plus accountability, the accumulated public record that lets a court, a client, or a promotion committee pin responsibility on a person. AI feedback can compress skill. It cannot mint a track record. If the binding scarcity was always accountable experience rather than raw competence, the inversion produces brilliant 26-year-olds who still cannot fill senior seats, and the gap survives in its institutional form after being solved in its cognitive form. Second, dense AI feedback exists exactly [where answers are cheap to check](../06-uncertainties/learned-verification.html): code that runs, proofs that verify, translations that parse. Those are the domains automating first, where a compressed apprenticeship is least needed. In the taste-and-judgment domains where seniors are scarcest, learned feedback is weakest, for the same reason those domains automate last. The inversion is real where it is least valuable and speculative where it is most valuable. Still. Every other response to the gap requires an institution to move against its own incentives. This one just requires somebody to build the tutor and a profession to accept its graduates. If you want the most valuable unbuilt thing in the entire labor story, it is this. ## What would prove this post wrong[#](#what-would-prove-this-post-wrong) The nice thing about a slow-motion problem is that it publishes its own scorecard. - The Recovery Test resolves around 2027 to 2028. If entry-level hiring recovers within a couple of quarters of aggregate white-collar hiring, the decline was mostly cyclical, and the corpus's 50% substitution estimate gets revised hard downward. - If entry-to-senior posting ratios stabilize through 2029, measured per profession rather than as a blended average, the institutional response is working and the commons failure becomes a lagged adjustment. One caveat guards the indicator: "entry-level" can stabilize because mid-level work got rebadged with a junior title. Check the seniority mix underneath, and check whether time-to-promotion is quietly lengthening. - If senior-to-junior wage premiums in exposed professions fail to widen by the early 2030s, the cobweb model itself was wrong and something stranger is going on. - Watch for firms or professional bodies selling explicit apprenticeship or associateship programs. That is the private form of the residency response, and its appearance is bullish even while posting data stays noisy. The forecast this post leans on says the ratios keep falling, the wage gap widens, and the serious institutional response starts only after the shortage is felt, which is roughly a decade too late. The saplings were always annoying. They shade nothing, they take up ground, and for years they produce nothing a client would pay for. A forest that optimizes them away runs beautifully for twenty years, and every quiet year is one more datapoint for the view that the worriers were wrong. The canopy is full. Nobody is counting saplings. ## Where this comes from Every number above is carried by a page in the corpus. These are the ones doing the work: [Game 4 - Labor](../02-games/4-labor.html)02-games/4-labor.md[Uncertainty 3 - The shape of the apprenticeship gap after institutions respond](../06-uncertainties/apprenticeship-gap.html)06-uncertainties/apprenticeship-gap.md[Law](../03-domains/cognitive/law.html)03-domains/cognitive/law.md[Education](../03-domains/cognitive/education.html)03-domains/cognitive/education.md[Diffusion - Labor and Institutions (B1, B6, B10, B11)](../07-indicators/diffusion/labor.html)07-indicators/diffusion/labor.md Or interrogate the whole thing directly: [ask the corpus](../ask/).