Somewhere right now, a first-year medical resident is learning to read a chest X-ray. She looks at the gray gradients, the branching vessels, the soft densities that might mean something or nothing, and she builds — slowly, through thousands of repetitions over years — a kind of intuition that will eventually let her see what others miss. That process has been essentially unchanged since the first teaching hospitals. It is about to change so completely that the people entering training a decade from now will experience something categorically different.
Not because AI will read the X-rays for them. Because AI will sit alongside them through every single case and teach them the way a brilliant, infinitely patient senior physician never could — annotating reasoning in real time, flagging the subtle density in the left lower lobe they just glossed over, asking them why they weighted that finding the way they did, and tracking across thousands of cases exactly where their perceptual model is drifting. The apprenticeship is becoming continuous, adaptive, and deeply personalized in a way that classroom instruction and occasional attending supervision simply cannot be.
This is the transformation that doesn’t get enough attention amid all the discussion of AI as a practitioner. The more profound near-term shift may be AI as a trainer of practitioners — an entity that knows the full literature, has seen the full distribution of cases, and can meet a learner exactly where they are.
Think about what that actually means across professions. A young structural engineer learns to develop intuition about load paths through years of checking calculations and occasionally seeing failures. With an AI collaborator that can run thousands of parametric variations in the background, explain the structural reasoning in plain language, and show exactly where a particular design assumption would become dangerous under edge conditions, that intuition can be built faster, deeper, and against a richer set of scenarios than any mentorship tradition has ever managed. The same logic applies to law, to materials science, to financial risk analysis, to the kind of judgment-heavy work that has always required years of accumulated pattern recognition.
What makes this plausible now, rather than just aspirational, is the convergence of a few things happening together. Long-context models that can hold an entire professional domain in working memory. Systems like the emerging class of reasoning models that can articulate not just answers but the structure of the reasoning behind them. And the growing body of work on personalized learning systems that track epistemic state — not just what a learner got right, but where in their reasoning chain they went wrong and why.
The generation entering professional training today will be the last to go through the old model largely intact. They will still have residencies and clerkships and bar exams and site visits. But they will also have, woven through all of it, a collaborator that never tires of explaining one more time, that has read every case study and every failure report, and that is paying attention to them specifically.
The generation after them will never know what it was like to learn without that. Their intuitions will be built faster. Their blind spots will be caught earlier. The professions they enter will look the same from the outside — there will still be surgeons and engineers and lawyers — but the cognitive baseline those people carry into their first decade of work will be qualitatively different from anything in the historical record.
That is not a small thing. Expertise has always been the great bottleneck — slow to create, impossible to scale, heartbreakingly unequal in its distribution. We are building the infrastructure to make it otherwise. The residents entering training ten years from now will be the evidence.