AI Isn't Crushing 'the Dev', It's Wiping Out the Interchangeable Junior (the Market Numbers)

Since late 2022, employment for developers aged 22 to 25 has fallen by nearly 20% in the United States, according to the Stanford Digital Economy Lab. Within the same companies, on the same roles exposed to AI, developers aged 35 to 49 saw their employment grow by +6% to +12%. This isn't an anecdote: it's a structural scissor.

The prevailing narrative keeps repeating, "AI is going to replace developers." I think that sentence is wrong, or rather that it's too vague to be useful. AI doesn't replace "the dev." It wipes out the commodity layer (the interchangeable junior, the cheap hands) and it rewards the augmented layer (the senior who orchestrates AI with clear specs and an industrialized production system).

The scissor: juniors drop, seniors climb

Three independent sources converge, and I spent time cross-checking them to separate the signal from the LinkedIn noise.

Is AI going to replace developers?

No. But it is violently reshuffling the deck by experience level.

The Canaries in the Coal Mine report from the Stanford Digital Economy Lab documented a drop of nearly 20% in employment for software developers aged 22 to 25 between late 2022 and mid-2025, based on real payroll data from millions of American workers. This decline is concentrated, for the first time, on a single age bracket. The 30-50 year olds aren't moving, or they're growing.

In France, the picture is similar. According to an economic outlook note from Insee, salaried employment in IT activities fell by 3% between late 2023 and late 2025. The contribution of 15-29 year olds (excluding work-study) accounts for -3.8 points of that decline. In Q4 2025, employment of under-30s in IT dropped by -7.4% year over year, reports Blog du Modérateur.

The market didn't stop hiring developers. It stopped hiring their junior versions.

Apec forecasts 61,160 IT manager hires in 2026, up from before. But profiles with less than six years of experience are growing by only 1%, according to data relayed by Maddyness. The PwC AI Jobs Barometer confirms the other side of the scissor: jobs requiring an AI skill are growing 7.5% while the overall market shrinks, with a 56% AI wage premium.

Why AI targets the commodity layer

The scissor isn't a cyclical accident. It stems from a technical reality I observe daily on assignments.

What makes a junior replaceable by AI?

The typical tasks of a first developer job (writing CRUD, wiring up REST endpoints, building a React form, writing unit tests, fixing surface-level bugs) are exactly the ones where Claude Code, Cursor, or Copilot excel. According to GitHub's 2023 study, 46% of code is already AI-generated in some projects, and 88% of developers who use it report being more productive, as Ingetis points out.

A subscription to an AI coding tool costs between 20 and 200 € per month. A junior on a permanent contract represents 35,000 to 50,000 € in annual employer cost. Some startups have made their call. Journal du Net quotes Arthur Huppert, CEO of Mia: "Artificial intelligence is better than a fresh graduate." Thomas Essig, lead developer at Mia, adds: "In France, we train generalist developers. You have to train them in-house for 4 to 6 years before they're operational."

Which developers does AI reward?

Those who possess what the philosopher Michael Polanyi called tacit knowledge: the ability to sense that an architecture is going to crack before it cracks, to break a project into testable blocks, to write specs that AI can execute correctly. The sociologist Yann Ferguson (LaborIA, Inria) put this concept back at the center of the debate on Maddyness: AI doesn't replace the junior, it replaces the explicit, articulable fraction of their work.

The fraction that can't be made explicit (judgment, prioritization, architecture choices, technical negotiation with the product owner) stays human. And it requires experience. On my assignments, I see the difference every day: an augmented developer with at least 8 years of experience who drives Claude Code with CLAUDE.md and ARCHITECTURE.md files and per-block acceptance criteria delivers faster than a team of three juniors, with controlled technical debt.

"AI doesn't kill the developer. It kills the job where someone was paid to type code that the machine types better."

Vincent Roye, June 2026

What this changes for companies that hire

If you're a CTO or founder, the question is no longer "to hire or not to hire." It's: hire what, and in what format.

Should you hire juniors or augmented seniors?

I have a sharp take: the answer depends on the stage of your product.

If you're building an MVP or a critical feature, you need immediate velocity. Not the time to train a junior for 4 to 6 months before they're autonomous. An augmented senior on a staffing basis at $210/day delivers a first commit in under 7 days. Over 12 months, the math often favors staffing over a permanent hire once you factor in the hidden costs (payroll taxes, management, onboarding, turnover).

If you're building a long-term team, keeping a junior pipeline makes sense, but only as a training investment. Not as production labor. Yann Lechelle's op-ed on Maddyness frames the right diagnosis: if companies cut off the junior entry pipeline today, they'll be short of seniors capable of supervising AI five years from now.

The trap is hiring juniors instead of seniors for budget reasons. I've seen projects where code generated by Cursor without senior review created more refactoring than the time it "saved." A junior without supervision doesn't produce "more slowly." They produce debt.

How do you measure the real velocity of an augmented dev?

The right metric isn't lines of code. It's qualified PR throughput: how many pull requests pass review and don't generate a production bug. An augmented senior ships 8 to 12 PRs per week on a well-scoped project. A junior with AI might ship the same volume, but with a 40% to 60% review bounce-back rate.

How a junior can make it in 2026

The picture is grim, but there's a way out.

Is it still worth becoming a developer in 2026?

Yes, but not the same kind of developer. The traditional path (bootcamp, internship, first permanent job on CRUD, gradual ramp-up) is closing, because the bottom brick has been commoditized. The path that stays open is the orchestrator's.

Concretely, a junior who wants to survive must:

1. Master scoped AI, not be subjected to it. Learn to write executable specs, acceptance criteria, project context files. Not just "ask ChatGPT to code."
2. Pick a deep technical specialty rather than staying a generalist. The junior generalist is exactly the profile AI replaces. The specialist (infra, security, data pipeline, native mobile) keeps an edge.
3. Document their projects publicly. As the AWS Developers discussion points out, a personal technical blog and an active GitHub remain the best signals for a recruiter, because they prove the ability to explain and build, not just to prompt.
4. Aim for seniority in 3 years, not 6. AI accelerates learning as much as it accelerates code. A junior who uses Claude Code to understand a complex codebase, who reads papers with NotebookLM, who debates architecture choices with the AI, levels up faster than their predecessors.

A junior's salvation is to stop being interchangeable as fast as possible.

Erick Wendel observes the same thing: his brothers, in training in 2026, are learning Flutterflow, N8N, database modeling with a UI, and building a working app in their very first semester. The traditional curriculum (algorithms, C++, pure theory for two years) is out of sync with the market.

The real risk: cutting the pipeline and paying later

The systemic danger, beyond the individual fate of juniors: the delayed effect on the entire industry.

What happens if no one trains juniors?

Five years from now, today's seniors will leave or change roles. If no junior has been trained between 2024 and 2028, who will supervise the AI in 2030? Who will have the tacit knowledge needed to spot that a hallucinating agent injected a security flaw into a critical microservice?

Victor, a developer on the O Novo Programador channel, offers a precise technical diagnosis: an average UI-creation task burns 200,000 tokens. When the harness compacts the context between sessions, it loses information. AI needs a human who knows what they want to build, not a human who presses "Enter."

My advice: don't cut junior hiring entirely. Reframe it. Train a junior alongside an augmented senior, not alone on a project with Cursor. For immediate velocity, bring in an augmented senior on a staffing basis. That's the combination that holds.