The seniors of 2034 are not being hired this year

Software development job postings in the United States ran 69.3 percent senior-level in the first quarter of this year. Entry-level postings were 4.5 percent. By June, Stanford's Digital Economy Lab had employment for 22 to 25 year olds in the most AI-exposed occupations running about 19 percent below where it would sit had it kept pace with their less-exposed peers since late 2022.

Both numbers get filed under jobs. They belong under training. The junior engineer was never bought for output. The output was the cheaper half of the deal, and the expensive half was a person who in four years would know things nobody had written down.

That half is being cancelled by accident.

The decline is real and narrower than the headline

The line you hear at conferences, that entry-level tech hiring has fallen somewhere between 30 and 50 percent, is not a measurement. It is four measurements averaged by people who did not read any of them.

SignalFire's 2026 State of Talent report puts new graduate hiring down roughly 65 percent at what it calls the Tech Majors, twelve firms including Alphabet, Meta, Apple, Amazon, Microsoft, Netflix, NVIDIA, Tesla, Uber, Airbnb, Block and Stripe, and down about 76 percent at early-stage startups. Both are measured against 2019. That is profile data drawn across roughly 650 million individuals rather than payroll, and 2019 is a generous baseline for any decline you want to demonstrate.

Indeed's Hiring Lab published a cleaner series in July. Entry-level postings were down 7.5 percent year on year in May, while senior postings rose 14.7 percent. In software development the split is close to absurd: 4.5 percent of first-quarter postings entry-level against 69.3 percent senior. The longer view complicates my own argument, so here it is. Between 2019 and 2025 the entry-level share of tech postings fell by only about one percentage point, while the mid-level share dropped 7.7 points and senior gained nine. Tech hollowed out in the middle before it closed at the bottom.

The Stanford figure is the one worth reading properly, because almost nobody does. The 19 percent is a relative gap, not a collapse. In absolute terms, employment of 22 to 25 year olds in the two most AI-exposed occupation quintiles fell about 11 percent between November 2022 and June 2026, while the same age group in the three least exposed quintiles grew about 10 percent. Young software developers are the sharpest case in the whole series.

Brynjolfsson and his co-authors are more careful than their coverage. They describe these as descriptive patterns rather than causal estimates. The gap narrows once you control for education. Their ADP sample produces larger effects than national survey benchmarks do. November 2022 also happens to be the top of a hiring bubble that was going to deflate whether or not anyone shipped a language model, and rates moved hard in the same window. Anyone selling you a clean causal story is selling.

The automation target and the training set are the same set. That is new.

Judgment is a by-product of boring work

The most useful thing I know about industrial temperature sensors did not come from a datasheet. It came from a night shift on a line where a reading drifted in a way no physical process could produce, and from a senior engineer who watched me chase it for six hours and refused to tell me the answer. A cable had been rerouted during a maintenance window and now ran parallel to a motor drive for about two metres. Electrical noise. Nobody has ever had to tell me twice.

That night cost the company a shift of my useless time. It also bought them an engineer who, years later, can read a fluent and entirely plausible diagnosis from a model and feel a specific itch that says check the physical layer first. Ask an assistant the same question today and you get four ranked causes in nine seconds, noise among them, with a code sample. Better outcome that week. Worse engineer that decade.

This is measurable, and the measurement is uncomfortable. METR ran a randomised trial in 2025 with sixteen experienced open-source developers across 246 real tasks in repositories they maintained, averaging five years of familiarity each. Tasks where AI tools were allowed took 19 percent longer. Afterwards, the same developers estimated the tools had made them 20 percent faster. Thirty-nine points between what happened and what it felt like, in about the most experienced population you could assemble.

Calibration is the product a senior actually sells. It is manufactured out of several thousand small occasions of being wrong in a way you could trace afterwards. There is no compressed version. You cannot read it and you certainly cannot prompt for it, because knowing which of the four ranked causes is worth checking first is the exact skill the ranked list was meant to replace.

The Stanford group found one result that should be read as a design brief rather than a forecast. Entry-level employment fell where AI applications automate work and barely moved where they augment it. Somebody chooses which of those to build. That choice is currently being made quarter by quarter by people who will not be in the job in 2034.

Shenzhen has the same problem with 12.7 million graduates

China will graduate 12.7 million people from its regular higher education institutions this year, 480,000 more than last year. Campus recruitment season, 校招, is genuinely expanding on the AI side. That is not the interesting part.

Caixin Weekly ran a cover story on 9 May that put it more bluntly than anything I have read in English: there is no longer any position completely unrelated to AI, and applicants are required to understand it and use it. The strong offers described in that piece go to graduates who turn up already carrying substantial internship experience.

So the gate moved. Entry now requires the experience that entry was supposed to provide, and the internship that supplies it is itself a junior rung, rationed by who can afford several months of near-unpaid work in one of the most expensive cities in China.

A 24-year-old told me over coffee near Shenzhen Bay in June that his offer came down entirely to six months of that, which he could take because his parents covered his rent. He was not complaining. He stated it the way you state a hardware requirement.

In Lagos the first job is also the passport

Nigeria's ICT sector grew 10.98 percent year on year in the first quarter of 2026, faster than any other sector, in its thirty-third consecutive quarter of growth. Telecommunications and information services alone reached 9.19 percent of GDP, up from 7.67. Demand is not the constraint here.

In San Francisco a junior role is income and training. In Enugu it is also a document. It is the first line on a record that a hiring system in Berlin or Singapore can actually read, the first reference a foreign employer will pick up the phone for, the first payment rail with your own name attached. Removing the junior rung removes a salary. It also removes the only reliable machine for converting a Nigerian degree into a credential the world outside Nigeria knows how to price.

A developer in Lagos showed me his GitHub in March. Eleven projects, several of them genuinely good. I asked why he had put a cache in front of one particular query and he could not tell me, because he had not decided it. He is 23. The tool that would once have taught him that decision made it for him instead, and the portfolio he was told to build as proof of competence has been quietly devalued by the same tool. Neither half of that is his fault.

The ATM argument, and where it breaks

The strongest objection is historical and it is a good one. James Bessen's account of bank tellers is the standard citation. American teller numbers rose from roughly 300,000 in 1970 to well over 600,000 by the early 2000s, straight through the ATM rollout that was supposed to end them. The machines cut tellers per branch from about twenty to thirteen between 1988 and 2004, which made branches cheap enough to multiply, and urban branch counts rose 43 percent. The work shifted from counting cash to selling relationships. The ladder rebuilt itself in a different shape.

Every automation wave since the power loom has produced this same prediction of collapse, and the totals have never behaved the way the prediction required. I am not going to pretend otherwise.

Three things get left out of the teller parable. First, the teller lost. The US Bureau of Labor Statistics now projects teller employment falling 13 percent between 2024 and 2034, with all 29,800 projected annual openings coming from replacement rather than growth. The ATM did not kill the job. The smartphone did, about thirty years later. "The ladder rebuilds" is true and useless without "on what schedule".

Schedules are the whole argument. A market can regrow demand for juniors in three years. It cannot regrow a 36-year-old with twelve years of judgment in three years, because the input to that is twelve years and there is no way to buy them in a hurry. Each year the intake stays shut lands undiminished on the senior supply of the mid-2030s, and no price signal arrives early enough to prevent it.

Then the structural point, which the analogy does not cover. ATMs took the cash drawer and left the conversation, so tellers went on learning the bank by working in the bank. Coding assistants and agents are aimed precisely at the bundle that constituted the apprenticeship: the routine ticket, the first-pass review, the reconciliation, the test nobody wanted to write. The automation target and the training set are the same set. That is new.

Where I could be wrong. The mechanism might be mostly cyclical, a correction to pandemic overhiring plus a rate shock, with AI as the story firms attach to a decision they were making anyway. New rungs are also appearing, and I see them: evaluation work, domain-specific data curation and agent supervision are junior work under different names. My worry is how they are being priced, as contract and piece work rather than employment with a path attached. That is an observation from where I sit rather than a dataset, and I would rather be argued out of it.

An accounting decision before it is a technology verdict

Training a junior was always slightly irrational for the individual firm, because a decent share of the return goes to whoever employs them next. It survived on three legs. Seniors were scarce, the convention was strong and juniors produced enough real output to cover their own cost. Only the third leg was ever load-bearing in a spreadsheet, and that is the one that went.

Which is why the steepest declines in the SignalFire data sit at the two extremes, the twelve largest firms and the early-stage startups. Both measure everything. Neither carries any slack.

The entry requirement now asks for what entry used to produce

Every posting I read this year, Chinese or English, asks for judgment about model output. Know when it is wrong. Know what to check before you believe it. Know what not to ship. That is senior work written into a junior requisition, and it used to be the residue of five years spent doing the boring version of the job.

The people signing these hiring freezes were juniors once, every one of them. They are pricing at zero the rung they are standing on, because from up there it has stopped being visible. The shortage will not announce itself. It will show up around 2034 as a company that cannot find anyone qualified to check what the model just shipped, and the fix will take exactly as long as it always did.

Sources

Stanford Digital Economy Lab, AI employment gap for young workers widens to 19% (August 2026) : https://digitaleconomy.stanford.edu/news/canariesaug26/

SignalFire, State of Talent Report 2026 : https://www.signalfire.com/blog/signalfire-state-of-talent-report-2026

Indeed Hiring Lab, The Labor Market Is Tilting Toward Seniority (23 July 2026) : https://hiringlab.indeed.com/2026/07/23/the-labor-market-is-tilting-toward-seniority/

METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity : https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/

James Bessen, Toil and Technology, IMF Finance & Development (March 2015) : https://www.imf.org/external/pubs/ft/fandd/2015/03/bessen.htm

US Bureau of Labor Statistics, Occupational Outlook Handbook: Tellers : https://www.bls.gov/ooh/office-and-administrative-support/tellers.htm

South China Morning Post, China braces for record 12.7 million graduates in 2026 : https://www.scmp.com/economy/china-economy/article/3333652/china-braces-record-127-million-graduates-entering-tight-job-market-2026

Techeconomy, ICT retains crown as Nigeria's fastest-growing sector with 10.98% expansion in Q1 2026 : https://techeconomy.ng/ict-retains-crown-as-nigerias-fastest-growing-sector-with-10-98-expansion-in-q1-2026/