The junior-dev cliff is real — but it's a training-pipeline problem, not a layoff story
The scary chart going around is real. Employment for 22-to-25-year-old software developers in the most AI-exposed work has fallen sharply while their older colleagues held steady. But "AI is taking junior jobs" is the wrong lesson to draw from it, and the right one is slower, less dramatic, and more worrying: we may be quietly switching off the machine that turns juniors into seniors.
What the data actually says
Stanford's Digital Economy Lab has been tracking this with real payroll data (their "canaries in the coal mine" work). As of their August 2026 update, the AI employment gap for young workers in exposed occupations widened to 19% — meaning employment for 22-to-25-year-olds in highly AI-exposed jobs is about 19% below where it would be had it kept pace with similarly-aged workers in less-exposed jobs. It was 15% a year earlier. For software developers specifically, employment among 22-to-25-year-olds is down roughly 20% since its late-2022 peak.
Two honesty notes, because the number gets mangled. It's a relative gap versus less-exposed peers, not a raw 20% headcount cut. And the "seniors are booming" half is overstated: the data shows older developers roughly flat, not surging. What actually grew ~10% was young workers in the least-exposed occupations. So the honest shape is: entry-level, AI-exposed roles took the hit; everyone else more or less held.
The reframe: a training deficit
Here's the part that isn't in the chart, and that I find more concerning than any single year's numbers. Junior work — the small, well-specified, forgiving tasks — is exactly what AI now does cheaply. So the rational move for a team under pressure is to hand those tasks to the agent instead of the new hire. Individually sensible. Collectively, a slow disaster, because those tasks were how juniors became seniors. The boring ticket, the code review you learn from, the "go figure out why this is slow" — that's the apprenticeship. If AI absorbs the bottom rung, the ladder doesn't get shorter; it loses its bottom rung, and in five to ten years the industry discovers it hasn't produced the senior engineers it now depends on more than ever (someone has to review all that AI output).
To be clear about what's evidence and what's argument: the decline is measured data. The pipeline-collapse consequence is a projection — a well-argued one (researchers at Bath's Institute for Policy Research have made it directly), but a forecast, not a fact. I'm flagging that because it's exactly the kind of claim that's easy to state too confidently.
If AI does the junior work better than a junior, the junior never does it, and never becomes the senior who was supposed to supervise the AI. We may be optimising away our own succession plan.
Don't let it be mono-causal
It would be dishonest to pin this all on AI, and the Stanford team is careful here too. The 2021-22 era of zero interest rates drove speculative over-hiring, especially of juniors; then came the fastest rate hikes in decades and a correction concentrated in exactly those over-created roles. A US tax change (Section 174) made capitalising software-developer salaries more expensive right in this window. And Stanford's own analysis found the AI-attributable decline is statistically clear only after 2024 — the earlier dips look like the ZIRP hangover, not the robots. So the honest read is multi-causal: rates and over-hiring explain a lot of the broad slump, and AI exposure explains why the entry level of certain jobs got hit disproportionately.
Why I care, as someone on the other side of it
I've been doing this since Drupal 6. I got to be a senior the slow way, by being handed work that was slightly beyond me and being allowed to struggle through it with someone watching. I use AI agents daily now and they're genuinely good at the kind of task I cut my teeth on. Which means the uncomfortable question isn't "will AI replace juniors" — it's whether those of us who benefited from the apprenticeship are now, task by rational task, declining to offer it to the next lot. I don't have a tidy fix. But "delegate the boring work to the agent" and "grow the next generation of engineers" are quietly in tension now, and pretending they aren't is how we sleepwalk into a senior shortage we built ourselves.