The industry stopped training juniors and has not said what replaces it

The number that should worry the industry more than it does: entry-level developer hiring is down 67% since 2022, with employment for developers aged 22 to 25 down nearly 20% since late 2022.

The mechanism is not mysterious, and that is what makes it awkward. AI coding tools are unusually good at exactly the work that used to be a junior engineer's training ground — boilerplate, small well-specified bug fixes, documentation, test scaffolding. Companies that once hired five juniors now hire two mid-level engineers with AI tools.

On a one-year view that is straightforwardly efficient. On a five-year view somebody has to explain where mid-level engineers come from.

Why those tasks existed

Here is the part that gets lost when the work is framed as low-value.

Nobody ever assigned a junior a CRUD endpoint because the CRUD endpoint was hard. They assigned it because writing it teaches you where the codebase keeps things, what the team's conventions are, how the review process works, what breaks in production and why, and — most importantly — how to tell a good solution from one that merely runs.

The output was almost worthless. The learning was the point. It was an apprenticeship with a deliverable attached, and the deliverable was the excuse.

When you automate the deliverable, you do not automate the learning. You remove the occasion for it.

The judgement problem

This connects to something visible in the productivity research: AI tools help most in clean codebases with good tests, and their output requires evaluation.

Evaluation is a senior skill. Recognising that generated code is plausible but subtly wrong, that it handles the happy path and ignores the concurrent case, that it has quietly introduced an N+1 query — that judgement was built by writing bad versions of the same thing and having someone explain why they were bad.

So the industry has built a tool whose value depends on senior judgement, and paid for it by removing the pipeline that produces senior judgement. That is a coherent short-term trade and an obviously incoherent long-term one.

Where juniors still get hired

The pattern is informative. Enterprises, financial institutions, defence contractors and healthcare technology firms still hire juniors consistently, because their codebases and compliance requirements demand human judgement in places automation cannot reach.

In other words, the organisations that still train people are the ones where the work is too consequential to hand over. That is a slightly damning observation about everyone else.

The expectations have also shifted. Juniors are now expected to arrive "pre-trained" on skills that used to be taught on the job, fluent with AI tooling and able to ship complete features. Which is to say: the industry has stopped offering apprenticeships and started requiring that candidates have already completed one.

What a small studio can actually do

I am not going to pretend a small team can fix a structural problem. But the incentives here are less one-sided than they look.

Hiring a junior in 2026 is cheaper in one respect than it was: a motivated junior with good tooling reaches useful output faster than they did five years ago. The tools that removed the training tasks also compress the ramp.

What has to change is what you assign. The old model — give them the simple ticket — no longer transfers much, because the simple ticket is now a prompt. The version that works is to put them on real work with a senior reviewing closely, and make the review the teaching. That is more expensive in senior attention, which is precisely why fewer firms do it.

The other thing worth saying plainly is that a studio which trains people ends up with engineers who understand its systems deeply, and that is not a charitable act. It is the cheapest way to get senior engineers who are also a good fit, in a market where hiring them directly is expensive and uncertain.

The open question

Nobody knows how this resolves. It is possible that the definition of "junior" shifts, that the first two years become about system design and evaluation rather than implementation, and the pipeline reconstitutes in a different shape.

It is also possible that a cohort simply does not get trained, and in seven years there is a visible shortage of people who can do the judgement work that the tools require.

What is not credible is the position that this is fine and requires no thought. Every senior engineer currently reviewing AI output learned to do it by doing the work that AI now does. That was not free, and it has not been replaced with anything.