The Apprenticeship We’re Automating Away

The most fluent newcomer on a team I watched this year was also the one who understood the system the least, and it took me longer than it should have to see that those weren't two separate facts. They were the same fact wearing a disguise. He was fast because he never had to struggle, and he didn't understand the system because he never had to struggle. The speed and the shallowness came from the same place.

For as long as I have done this work, there was a hidden curriculum underneath the job, and almost no one called it by name. You arrived, and you were handed the small, unloved tasks: the flaky test, the bug three layers down, the corner of the codebase nobody wanted to touch. The work itself was rarely important. What mattered was that to finish it you had to wander — read code you did not write, follow a call you did not understand, get lost, ask, and slowly assemble in your head a map of how the whole thing actually held together. The difficulty was not in the way of the learning. The difficulty was the learning.

We told ourselves this was just paying dues, the unglamorous tax you settled before you got to do the real work. I no longer think that is what it was. The grunt work was an apprenticeship disguised as drudgery. It did not just produce a fixed bug; it produced a person who now carried a little more of the system inside them. The friction was the pedagogy. The slowness was the point.

AI removes the wandering, and this is the part I think we have not sat with honestly. A newcomer — a junior, or a senior who is simply new to your codebase — now describes the problem and receives a working answer, correct and complete, without ever traversing the ground between the question and the solution. The destination arrives; the journey is skipped. And because the map was only ever a byproduct of the journey, no map forms. They are productive almost immediately, which is exactly what hides the problem, because productivity was never the thing we should have been watching.

The cost is real, but it is deferred, which is the most dangerous kind of cost. It does not show up in the onboarding metrics or the first sprint. It shows up months later, the first time something fails in a way the model cannot account for, when the person closest to it reaches inside themselves for the mental model that would let them reason about it and finds the shelf empty. Multiply that by a team and you get an organisation that can produce code inside a system none of its people actually hold — where the understanding that used to live in humans now lives only in a tool, one that cannot be woken in a crisis and asked what it was thinking.

I learned to write the same way I learned to engineer, which is to say badly, slowly, and by doing the parts I wanted to skip. The first novel I finished was not good, and finishing it anyway was the entire apprenticeship; the book was the byproduct, the writer was the product. A model will hand you a competent paragraph now, the way it hands a newcomer a competent function, and in both cases the competence is real and the learning is absent. You cannot shortcut your way to judgment, because judgment is mostly the residue of having struggled through enough cases to feel the shape of the next one. Take the struggle away and you do not get a faster apprentice. You get someone who never served the apprenticeship at all.

I am not arguing for manufactured hardship, or for making people grind through work a machine could do, as if difficulty were a virtue on its own. I am arguing that we now have to do on purpose what friction used to do for free: protect some portion of the hard, unassisted traversal as the way people learn the system, not merely the way they feed it. The alternative is quiet, and it arrives dressed as progress — a generation of engineers fluent in answers and strangers to the systems they are paid to keep. The cheapest onboarding we have ever built may be teaching people the most expensive thing of all: how to operate something they were never required to understand.