The junior developer collapse is real, but misdiagnosed

The headline everyone's passing around is that AI torched the market for junior programmers. It's true. It's also the least interesting thing you can say about the numbers, because it collapses two very different phenomena into one scary sentence. One is a genuine wipeout of early-career hiring. The other is a quiet reclassification of what "writing code" is even worth. If you're early in your career, or you hire people who are, the difference between those two decides what you should actually do next.

So let's separate them.

The carnage is concentrated, not broad

Start with the number that launched a thousand LinkedIn posts. Using ADP payroll data, Stanford's Digital Economy Lab finds US software developers aged 22 to 25 are down roughly 19% from their late-2022 peak. Every cohort over 30 grew over the same window, with 41-to-49-year-olds up 14%. After controlling for firm-level shocks, the relative decline for young workers in AI-exposed jobs still lands around 16%, and it concentrates specifically where AI automates work rather than augments it.

Now the part that gets dropped: in aggregate, developer employment is fine. BLS counts software developers rising from 1.53 million in May 2022 to 1.69 million in May 2025, up about 10% straight through the AI era. Total developer employment is up 4.4% since October 2022. Computer and mathematical occupations grew 1.3% from May 2024 to May 2025, faster than the economy's 0.8%. A Danish study using government payroll records can rule out any aggregate AI employment effect bigger than about 1%.

Both things are true because juniors (defined here by age, which is a real caveat) are only about 8% of the developer workforce. A catastrophe for 8% barely moves the average. That's why every study that looks at averages finds nothing and every study that looks at juniors finds a crater. They're reading different slices of the same file.

Worth noting the confounders honestly: the same period brought the ZIRP unwind, the Section 174 R&D tax change, and a post-pandemic hiring hangover, and companies themselves attributed only about 4.5% of 2025's announced layoffs to AI. But none of those explain why the damage is so precisely aimed at 22-to-25-year-olds in automatable roles while their 40-year-old colleagues thrive. Given tech's well-documented ageism, you'd expect the opposite if the market were just generally rough.

The title is dying, not the work

Here's the mechanism the panic misses. Break the BLS occupation categories apart and you can see exactly which kind of job AI is eating.

xychart-beta
    title "US occupation change, May 2024 to May 2025 (%)"
    x-axis ["Programmer", "Web dev", "QA", "Software dev", "Sys analyst", "Data sci"]
    y-axis "Percent change" -20 --> 15
    bar [-16, -11, -6.5, 2, 4.4, 12]

The "computer programmer" category, BLS's label for people who write code to someone else's specification, fell 16% in a single year against a projected decline of 6% per decade. Web developers fell 11%, QA testers 6.5%. Meanwhile data scientists grew 12%, systems analysts 4.4%, and the broad "software developer" bucket grew 2%.

See the pattern? The jobs shrinking are the ones where the deliverable is code written to spec. The jobs growing are the ones where the deliverable is judgment about what code should exist. That's the whole story in one chart. This also lines up with the timeline: the junior line didn't fall off a cliff when ChatGPT launched in late 2022. It drifted, then deteriorated fastest in 2024 and early 2025, exactly when coding assistants stopped autocompleting lines and started closing tickets. Agentic tooling turned up the heat, not the chatbot.

And the flood of new entrants that abstraction layers always produce? It showed up. GitHub added 36 million new accounts in its last Octoverse year, its fastest ever, more than one new developer per second, plus 121 million new repositories. Eighty percent of those arrivals used Copilot within their first week. The new programmers are real and measurable. They just don't call themselves programmers, and they're not competing for the entry-level rec you posted.

The pipeline math is a slow-motion own goal

Executives know this, at least out loud. AWS's Matt Garman called cutting juniors in favor of AI "one of the dumbest things I've ever heard," asking how it works when ten years out you have nobody who's learned anything. GitHub's former CEO Thomas Dohmke called the same idea "overblown." Sundar Pichai frames AI as an accelerator, and Satya Nadella, even while noting AI now writes 20 to 30% of Microsoft's internal code, insists juniors still need the fundamentals.

Then look at behavior. Salesforce said it stopped hiring new software engineers in 2025. Klarna froze developer hiring back in late 2023 (and quietly rehired humans after that didn't pan out). Anthropic's Dario Amodei has put entry-level roles "in the crosshairs." The gap between the reassuring quotes and the payroll data is the actual signal.

The reporting doesn't even agree on how bad the posting collapse is. One read puts entry-level software postings down 28% from their 2022 peak; Indeed data via FRED shows listings down roughly 35% from pre-2020 and about 70% from the 2022 peak; another tally has entry-level postings dropping 60% between 2022 and 2024. Different measures, same direction, all ugly. Meanwhile CS grads now carry a 6.1% unemployment rate, higher than liberal arts, and workers aged 22 to 27 sit near 7.4%, roughly double the national rate. There's a reason 64% of Gen Z say they're worried about layoffs versus 45% of millennials.

The deeper problem is structural. Juniors have always been how you manufacture seniors. Skip that layer for a few years and you get a demographic notch in the talent supply that no amount of Copilot fixes in 2032. This is also happening while capital visibly rotates from headcount to hardware. In Seattle, downtown office vacancy has hit 37%, the worst of any major US metro, and analysts tie the shortfall directly to tech's pivot from spending on people to spending on AI data centers. The money didn't leave the industry. It moved from payroll to GPUs.

What to actually do if you're early-career

Stop optimizing to be a fast code-typist. That role is the one the machines are eating, and 84% of developers already report using AI tools in their workflow, so "I can implement a ticket" is now table stakes, not a differentiator.

Concretely:

• Aim for the judgment jobs, not the spec jobs. The growing categories (data science, systems analysis, the broad developer role) reward people who decide what should be built and why. Learn to read a fuzzy requirement, argue about tradeoffs, and own an outcome. That's the work agents are worst at.
• Get fluent as an operator of agents, not a competitor to them. Being able to drive an agentic workflow, review its output critically, and catch where it's confidently wrong is a distinct, valuable skill. The juniors who thrive are AI-native by default; use that.
• Build the fundamentals the AI hides. Nadella's "think computationally" point is self-serving but not wrong. When the abstraction leaks (and it always leaks), the person who understands what's underneath is the one who ships.
• Consider the side door. The 36-million-account wave means a lot of software is now getting written by people whose title isn't "developer." Domain experts who can also build are in a strong spot. You may enter the field through a role that doesn't say "engineer" on it at all.

If you hire, the case for keeping a junior bench is boring and correct: it's cheap relative to the ten-year cost of a talent gap, and the people you train on AI tools now become the seniors who can actually supervise them later.

The scary headline gets the fact right and the lesson wrong. AI didn't kill programming. It killed the commodity version of it, the part where you turn someone else's spec into syntax. That was always the least defensible thing a human did in this job. The uncomfortable news for new grads is that it was also the traditional on-ramp. The rung didn't move. It got removed, and nobody's agreed yet on what replaces it.