The Missing Rungs

Stanford AI Index Report is interesting.

Employment for software developers aged 22 to 25 has fallen nearly 20% from its 2024 peak.

Not 20% below projection. Not 20% below some theoretical optimum. Twenty percent fewer young people actually working as software engineers compared to two years ago.

This is not a prediction or a warning.

It is a measurement, derived from ADP payroll records covering 3.5 to 5 million workers, cross-validated by researchers at Stanford’s Digital Economy Lab.

The same dataset shows that software engineers aged 35 to 49, the experienced ones, saw 6 to 9% employment growth over the same period.

So AI is not eliminating software engineering. It is inverting the pyramid: compressing the bottom, expanding the middle and top, and erasing the entry-level tier that made the whole thing work.

This matters in a way that is distinct from the usual job loss anxiety, and I want to explain why.

What the Junior Role Actually Was

If you’ve worked in a large software organization, you know what junior developers actually do. They don’t design systems. They don’t make architectural decisions. They write unit tests, fix small bugs, implement well-specified features, do code review on other junior developers’ work,

And

Most importantly, they make mistakes in low-stakes situations where senior engineers can catch and correct them.

That last part is the hidden function.

Junior roles are not primarily about output. They are primarily about training. You learn to be a senior engineer by being a junior engineer for several years, getting your mistakes corrected, absorbing patterns from code reviews, developing intuitions about what makes systems fail.

The output of a junior developer is often slower and messier than what an AI agent would produce. But the process is what makes the senior engineer fifteen years later.

AI can outperform a competent junior engineer on most of the measurable tasks a junior engineer is asked to do.

The raises an uncomfortable question though

Where do senior engineers come from in 2036?

The Apprenticeship Problem

I thought of checking to see if this is a really unique problem for this generation or have we seen this before.

I found a historical analogy.

In the early twentieth century, the (American) legal profession faced a similar structural shift. Before law schools became the standard path, lawyers were trained through apprenticeship. You read law in a senior attorney’s office, handled small tasks, observed how experienced practitioners managed cases, and gradually took on more responsibility as you earned it. When law schools took over, the profession did not immediately produce worse lawyers, it produced lawyers faster and at greater scale. But something was also lost: the direct transmission of practical judgment, the daily correction of mistakes in context, the mentorship relationship.

The legal profession survived this shift, though it took decades to develop new mechanisms for practical training (clinical programs, associate tracks at firms, public defender offices).

The analogy to software engineering is imperfect but instructive. Using AI for coding is like replacing the law firm associate track. The mechanism that converts classroom knowledge into practical expertise is being automated away at the same moment that capability is accelerating.

Companies are hiring fewer juniors, relying on AI to do what juniors used to do, and congratulating themselves on productivity gains. The Stanford data shows that 26% productivity gains in software development are real. But those productivity gains are accruing to workers who already have the expertise.

The pipeline that creates expertise is narrowing.

New graduates, juniors and the future

It is worth sitting with the specific Forrester projection here: a 20% decline in computer science enrollments as prospective students respond to deteriorating job market signals.

Makes sense.

If the job market for entry-level programmers is visibly contracting, fewer people should major in computer science.

But this has consequences.

The senior engineers who will be needed to guide AI systems in 2030 have to start somewhere in 2026. If the 2026 cohort doesn’t enter the field because the entry-level positions have been automated away, the 2030 shortage may be real and painful.

This is not an argument for protecting junior engineering jobs artificially. It is an argument for taking seriously the question of where expertise comes from and whether there are mechanisms, educational, organizational, or otherwise that can substitute for the apprenticeship function that junior roles provided.

Some possibilities are obvious: more clinical-style programs, structured mentorship, AI-assisted learning environments that preserve the feedback loop even when the commercial context is gone. None of these are being built at the scale the problem will require.

The AI Index tells us adoption is at 53% and rising. It tells us productivity gains are 14 to 26% on structured tasks. It tells us consumer surplus is $172 billion and growing. These numbers are genuinely impressive and the progress is real.

What it cannot tell us, because the data does not yet exist, is whether the mechanisms for developing the expertise that keeps these systems well-governed are keeping pace with the systems themselves.

On that question, the most honest answer is that we do not know.

And not knowing, in this case, is the most important thing to know.