Walk into any junior dev community in 2026 and the conversation has the same shape. They’re not getting hired. Or they get an internship and no contract after. Or they’re treated as adjuncts to senior people who don’t really need them anymore. The pipeline has stopped pulling them in. Everyone notices. Almost no one wants to say what it means.
I want to say it. And I want to say what we’re going to have to do about it, because I think we have about eighteen months before the current configuration cements into something we won’t be able to repair.
The wrong culprit
The reflex is to blame the AI. It’s a convenient culprit and it’s the one already in the room. But the AI is the amplifier in this story, not the cause. The cause is structural and it sits one layer up: the production of code stopped being the bottleneck about eighteen months ago, and we kept training, recruiting, and evaluating juniors as if it still was.
That’s the second half of an observation I started in The Reservoir Was Already There. The first half was that production stopped being the limiting factor for the senior engineers who chose to work with the new tools. The second half is the consequence: we have not adjusted the system that produces the people who would be doing the judgment work that’s now scarce.
When the bottleneck moves and the system feeding the bottleneck doesn’t move with it, the entry-level seats disappear first. That’s what’s happening to juniors. Not because AI is doing their job — because we never updated what their job was supposed to be.
Three blocks
The system that should have moved is blocked in three places at once, and none of them is going to unblock itself.
The first block is people. Specifically, the senior engineers who decided two years ago that the new tools weren’t for them, and never revisited that decision.
They’re not passive in this crisis. They’re active in it. They validate pull requests with reflexes from a practice they refuse to learn. They keep the seats they have without preparing anyone to take them. And when their organization asks them to mentor a junior, they pass down the model of engineering they understand — which is the model the market is in the process of retiring. This isn’t a moral failing on their part; it’s an unwillingness to do unpaid retraining for a transition they didn’t ask for. That’s understandable. It’s also incompatible with mentoring the people who will have to work in what comes next. You can’t teach a craft you’ve decided not to practice.
The second block is institutional, and it operates at two levels at once.
The first level is general education — primary, secondary. The level where children spend most of their cognitive formation. The back-to-school priorities announced this year for French national education are mastery of language, scientific reasoning, the school climate, and support for personnel. Every one of these is necessary. Every one of these is silent on the transformation that is already reshaping how children will interact with information for the rest of their lives. The official priorities do not contain the word AI. Not once. Twelve-year-olds in 2026 are using these tools every day, forming their cognitive habits unsupervised, and the system meant to teach them how to think has decided that this is not its problem yet. It will be its problem eventually — but by then the habits will already have formed, and the formation will have happened without any of the framing the institution was supposed to provide.
The second level is post-secondary — engineering schools, computer science programs, bootcamps. Every curriculum I’ve looked at in the last year still produces graduates trained for the role that’s disappearing. Syntax. Frameworks. Patterns. The capacity to produce code, evaluated on volume and conformity. Almost none of them teach what’s actually scarce now: the capacity to judge what a system produces. To refuse output that compiles but doesn’t fit. To articulate why one architecture is wrong even when no one can prove it yet. AI in those programs is either banned — which is unrealistic, so students learn to use it badly in private — or ignored, which is worse, because, again, they form their habits unsupervised.
The repetition is not accidental. The same failure mode runs through both levels of the education system, with different consequences at each end. At the post-secondary level, the result is a generation of graduates the market doesn’t want. At the primary level, it’s an entire cohort forming the cognitive habits that will define their adult relationship to information, with no one in the room to shape that formation. The second one is worse than the first, and almost no one is talking about it.
The third block is broader. No public policy. No transition framework. No coordinated effort between schools, employers, and public actors to map what the new entry-level role looks like. Consulting firms continue to bill juniors at day rates calibrated for a market that no longer values what those juniors are being asked to do. Public bodies continue to fund training programs that produce graduates the market is about to reject.
This third one is the hardest to attribute. There’s no specific actor refusing to act — there are a thousand actors, each of whom thinks it’s not their problem. The result is the standard outcome when no one’s problem is everyone’s: nothing.
What needs to happen
The blocks won’t unblock themselves, and they won’t unblock in the order one might hope. So the question is what each actor can do without waiting for the others.
Education has to redesign the curriculum at both levels. Primary and secondary need to integrate the new cognitive landscape into the basics — not as a separate “digital” subject taught once a week, but as a thread running through how language, reasoning, and evidence are taught. Children who learn to read in 2026 also need to learn to recognize what’s been generated for them, to ask whether the source can be trusted, to notice when a fluent paragraph is actually empty. That’s not an IT skill. It’s the new shape of literacy. Treating it as optional, or pretending it can wait until specialization, is a decision to outsource that skill formation to the platforms — which have their own incentives and are not the institutions we should want shaping how a generation thinks.
Post-secondary needs to rebuild what it teaches, not extend it. Adding an AI module to an existing program produces graduates who know one more tool. That’s not the change required. The change is to make judgment the central skill from the first week. AI in every course, but not as a productivity tool — as a material to critique. Evaluations that grade the ability to refuse, refactor, and explain — not just the ability to produce. Graduates whose first competency on a résumé is “I can tell you why this code is wrong” rather than “I know React.”
Organizations have to redefine the ladder, publicly. The expectations on the books today are mostly inherited from a world where a junior earned trust by producing a lot of acceptable code. That ladder no longer maps to anything real. The replacement isn’t subtle: a junior in 2026 who’s been well-trained should arrive able to operate the tools, reject their bad output, and explain to a senior why a given proposal is structurally wrong. That is the new entry-level skill. Organizations that can’t articulate this on a job posting are still hiring for a role that doesn’t exist, and they will keep failing to find people who fit it.
The senior engineers who mentor — and I want to be direct, this is a smaller group than the population that should — have to teach using the tools the junior will spend their career on. Not despite them. With them. The form of this work isn’t mysterious. Sit next to a junior. Watch them prompt. Stop them when the output is wrong. Ask them why it’s wrong. If they don’t know, walk them through it. If they know but accepted anyway, ask them why they accepted. Do this for two years and you produce engineers who can judge. Refuse to do this and you produce operators of prompts.
And then there’s the other group. The seniors who’ve decided the new practice isn’t for them.
You’re allowed to make that decision for your own work. You’re not allowed to make it for the people you’re supposed to be training. If you don’t practice with the tools, you can’t validate code produced with them, because you don’t have the calibration to know what’s good or bad in that output. You can’t mentor someone whose actual job is to operate in a practice you refuse to enter. The honest move is to step out of the evaluation and mentoring chain. Keep your own work. Don’t gatekeep the next generation’s.
I know this reads harsh. It’s the same standard we apply everywhere else: you don’t supervise work you can’t do. It’s only being applied unevenly here because the work changed faster than the chain of authority.
The window
We’ve seen this story before, more slowly. The previous generations had forty years to prepare for transitions everyone could see coming — demographic, fiscal, ecological. The whole time, we knew the systems needed updating. Nothing was done at the scale of what was needed. The cost was deferred to the generations that followed.
AI is the same dynamic, compressed from forty years to three. We have eighteen months, maybe twenty-four, before the current configuration cements. After that, we will have produced a generation trained for a role that no longer exists, and abandoned them to a market that no longer wants what we taught them.
Imagine three more years of this. We don’t have to imagine — we’ve seen the long version, and the long version got the same answer every time: nothing. The fast version is happening now. The answer is still up to us, but the window is small, and the people who will pay the cost if we get this wrong are not the people in the room making the decisions.
That last detail is the one I keep coming back to. The juniors who can’t find work in 2026 are not in the meetings where their training and their hiring criteria get decided. They will not be in those meetings in 2027 either. We are deciding, on their behalf and without consulting them, what kind of career is going to be available to them. The least we owe them is to make that decision with the new reality in view, instead of with the old one we found more comfortable.