Yes, the entry-level market has collapsed. That part isn't in dispute and the numbers are worse than most people realise. What is very much in dispute is why — and the best available answer is probably not the one you've been reading.
How bad is it actually?
SignalFire's talent data has entry-level hiring across the tech majors down 65% since 2019. Graduates of the top 20 US computer science programmes are about 45% less likely to take an engineering role at one of those companies than they were a few years ago. Turning junior developers into something a commercial team can run is the problem XenGrowth's work on go-to-market systems works on.
Stanford's Digital Economy Lab, working from ADP payroll records rather than surveys, finds employment for 22-to-25-year-olds in the most AI-exposed occupations sitting roughly 19% below where it would be had it tracked less-exposed peers. Experienced workers in the same occupations show no comparable gap.
And the New York Fed puts recent-graduate underemployment near 42% as of 2026:Q2 — the share working in jobs that don't typically require a degree. That number is more revealing than the unemployment rate, because it counts the graduates who found work, just not the work they trained for.
One honest caveat before going further: 2019-2021 was an unusually generous hiring market, inflated by cheap capital. Some of the 65% is a return from an anomaly rather than a fall below normal. Not all of it, though — not with a 42% underemployment rate attached. The XenGrowth resource library works through the operations side of this in more operational detail.
So is AI the cause?
This is where it gets interesting, and where almost every article on the subject stops before the interesting part.
Peter John Lambert at the LSE and Yannick Schindler at the Ellison Institute took Revelio Labs résumé data covering 243 million new hires and Lightcast data covering 407 million job postings, across the US, UK, Canada and Australia, 2017 to 2025. Then they did the thing that separates analysis from commentary: they put two candidate explanations in the same model.
When work-from-home exposure and AI exposure are included together, the AI coefficient attenuates sharply — often to statistically indistinguishable from zero. Work-from-home exposure remains a strong and robust predictor of declining junior share across every specification they tested.
Two variables moved at the same time, in the same industries, for the same five years. Almost everyone attributed the effect to the one that was more interesting to write about.
Their mechanism is supervision cost. Remote work makes monitoring, mentoring and incidental learning more expensive, and those costs land disproportionately on the people who need the most managerial investment. A junior hire is not a unit of output; they are a training investment that pays back over two years. Anything that raises the price of training reduces the quantity purchased. For the AI agents and marketing automation angle, see XenGrowth on AI agents and marketing automation.
I don't think this closes the question. Stanford's substitution/complement split is real and points at AI doing something specific. Correlational models with two collinear variables are exactly where confident conclusions go to die. But the LSE result should be enough to stop anyone treating AI displacement as settled, and it is conspicuously absent from the discourse.
Explanation | Evidence for it | Evidence against it |
|---|---|---|
AI automates junior work | Stanford's 19% gap concentrated in substitution-exposed occupations | AI coefficient collapses once WFH is controlled for (LSE, 243M hires) |
Remote work raised supervision cost | Robust predictor across every LSE specification, four countries | Doesn't explain why the gap tracks AI exposure specifically in Stanford's data |
Interest rates and budget discipline | The 2019-21 baseline was inflated by cheap capital; cuts hit training budgets first | Doesn't explain why the cuts fell so disproportionately on one cohort |
Experienced ICs preferred under uncertainty | SignalFire observes exactly this preference across company sizes | Describes the behaviour rather than explaining what caused it |
What exactly was the bottom rung made of?
It helps to be specific about what an entry-level engineering job actually consisted of, because the abstract version of this argument keeps sliding around. Junior work was, almost by definition, bounded problems with known answers, handed out so that a more experienced person could check the result. That structure was not an accident or a way of getting cheap labour. It was the training mechanism: you learn judgment by making small decisions where being wrong is survivable and somebody notices.
Which is exactly why the rung is vulnerable from two directions at once. A bounded problem with a known answer is the cheapest thing a model produces, so the tasks got cheaper. And a task whose entire value is the supervised feedback loop is worthless without the supervision, so remote work made the wrapper expensive. Both stories attack the same rung through different parts of it, which is probably why the statistics are so hard to separate.
What a junior actually did | Why it existed | What happened to it |
|---|---|---|
Implement a well-specified ticket | Safe practice at translating intent into code | Cheapest thing a code generator does |
Write tests for existing behaviour | Learn the codebase by describing it | Largely automated, and the automatable kind was always the low-value kind |
Fix small, well-triaged bugs | Learn debugging with the search space pre-narrowed | Triage itself is now partly automated, so the narrowed version arrives less often |
Sit near someone senior and absorb context | The actual mechanism of the apprenticeship | Removed by distributed work — the LSE paper's supervision-cost effect |
Get code reviewed line by line | Where taste is transmitted | Review capacity is now consumed by generated code volume |
Read that table and the 2030 problem becomes concrete. Every row in the middle column is a training function, not a production function. An organisation that deletes all five has not become more efficient; it has stopped manufacturing the seniors it will be bidding for in five years.
Why does the cause matter so much?
Because the two explanations have opposite implications, and one of them is actionable.
If AI genuinely automated entry-level work, then the rung is gone structurally and nothing a company does brings it back. That story is fatalistic, and it is also the one that lets every hiring manager off the hook.
If the binding constraint is supervision cost, the rung is not gone. It got expensive. Expensive things can be made cheaper deliberately: co-located onboarding periods, explicit apprenticeship structure, mentorship treated as a staffed responsibility rather than something senior engineers do in the gaps. None of that is exotic. It is what the industry did routinely before it stopped.
And there is a structural argument that the current equilibrium cannot hold. The BLS projects 15% growth to 2034, roughly 129,200 openings a year. Those openings need experienced engineers. Experienced engineers are manufactured out of junior ones over about five years. An industry that stops making them is drawing down a stock it isn't replacing, and it will notice around 2030. XenGrowth on AI search, GEO and discovery works through AI search, GEO and discovery in more operational detail.
What should you do if you're the one hiring?
The uncomfortable framing is that every individual decision not to hire a junior is defensible, and the aggregate of those decisions is a profession that stops reproducing itself. That is a textbook collective action problem, and the only companies that escape it are the ones that decide to on purpose.
Budget mentorship as staffed time, not goodwill. If the constraint is supervision cost, the fix is paying for supervision explicitly rather than hoping senior engineers absorb it between tickets
Front-load co-location. The LSE mechanism implies the first few months carry most of the learning value, which makes a short in-person onboarding period a far cheaper intervention than a full return-to-office mandate
Give juniors work with slow feedback, deliberately. The tasks with fast checkable answers are the ones that got automated; judgment only develops on problems where being wrong takes a while to surface
Count review capacity as a real constraint. Generated code volume is consuming the review bandwidth that used to transmit taste, and adding a junior to a team with no review headroom sets both of them up to fail
Notice that you are competing for a shrinking pool later. Hiring one junior a year is cheap insurance against bidding against everyone else for mid-level engineers in 2030
What should you do if you're the junior?
None of the above helps you this month, so here is the part that does. The consistent signal across all four datasets is that what got devalued is bounded, checkable, supervised work — and what did not is evidence that you can operate without supervision.
Optimise for evidence of unsupervised judgment, not credentials. The specific thing employers stopped wanting to buy is the training period. Anything that shortens it — a real system you built and operated, an open-source contribution you saw through review, an incident you diagnosed — attacks the actual objection
Take the in-person or hybrid role if you can, early. If the LSE mechanism is right, co-location is worth real money to you in learning rate, and it is worth it to them in supervision cost. That alignment is the strongest argument you have
Target companies where engineering is close to the product. Smaller organisations are where juniors still get scope, and scope is what converts into the judgment the market is short of
Learn to verify, not just to generate. Veracode's benchmark found 45% of generated code carrying an OWASP Top 10 flaw. A junior who can reliably catch that is solving a problem their employer measurably has
Do not wait for the market to normalise before starting. The 2030 shortage argument says the correction is coming, but it arrives on a five-year lag, and you need the five years of experience to be already accumulating when it does
The headline 'AI killed junior developer jobs' is clean, dramatic, and probably wrong on the best evidence available. The messier version — a hiring market correcting from an anomaly, a remote-work shift that quietly made apprenticeship expensive, and an AI effect that is real but smaller than advertised — is harder to compress into a sentence.
It also happens to be the version where somebody can do something about it.
Further reading from XenGrowth
The XenGrowth resource library — what you'll learn: how the commercial side of this work is run, across search, automation and revenue operations.
XenGrowth on AI agents and marketing automation — what you'll learn: how the teams who own AI agents and marketing automation plan and measure it.
XenGrowth on AI search, GEO and discovery — what you'll learn: how the teams who own AI search, GEO and discovery plan and measure it.
Where this work meets go-to-market
For the marketing and revenue operations view of junior developers, see XenGrowth.
Five questions on the studies behind this post. Most of the public argument runs on one number quoted without its controls, and the controls are where it gets interesting.
