Practice

The Generation We Will Not Have Trained

Automating junior work can preserve knowledge while breaking the mechanism that trained experts.

Here is a question almost no expert profession has ever asked itself, for the simple reason that it never had to: can a profession keep its knowledge while ceasing to produce the people capable of understanding it, contesting it and renewing it? The question seems abstract. It is becoming, under the effect of artificial intelligence, one of the most concrete and urgent there is. For all expert professions share a discreet mechanism of reproduction, never formalized, that AI precisely undoes.

This mechanism holds in one sentence, and this sentence is true for law as for medicine, aviation, architecture or engineering: every profession manufactures its experts by having beginners do the work the experts no longer do. The seasoned surgeon no longer opens, no longer sutures the elementary gestures; the resident does, and it is by doing it that they become a surgeon. The captain no longer fills in the routine checks; the co-pilot does, and it is by doing them that they learn to command. The partner no longer drafts the first version nor compares the clauses; the associate does, and it is by doing it that they learn to judge. Everywhere, the beginner inherits the work the expert has shed, and this work is exactly what turns them into an expert.

Every profession manufactures its experts by having beginners do the work the experts no longer do.

The mechanism no one had formalized

This mechanism has a remarkable property: it was never designed. No profession instituted it deliberately; it imposed itself, as a natural consequence of the division of labor between those who know and those who learn. Beginners were given the lowest tasks out of economy, because it would have been absurd to use experts for them, and training came as a bonus, unwilled and unnamed. The system manufactured its own experts as a side effect of its ordinary production, and this apparent gratuity masked a deep dependence: each profession reproduced itself from the bottom up, through the work all judged the least noble.

Because it was never designed, this mechanism is fragile in a particular way: it can be destroyed without anyone deciding to, simply by removing the tasks that carried it. Yet that is exactly what automation does. When a machine takes over the elementary suture, the routine check, the first research, the first draft, it renders an obvious service, and no one thinks to complain. But it removes, in the same gesture, the material support by which a beginner became an expert. The visible work is absorbed; the invisible training that lodged in it disappears with it, silently, without decision, without anything in the indicators signaling it.

Law faces this dynamic early, because AI there quickly automates the junior tasks: research, reviews, clause comparisons, first drafts, summaries. But it would be wrong to think the phenomenon proper to law. It lies in wait for every profession whose entry work AI learns to execute. Assisted medicine, assisted engineering, assisted consulting will meet the same erosion, only staggered in time. Law is not the exception; it is the outpost, the place where one can observe, in advance, what awaits all expert professions when the machine absorbs the work by which they reproduced themselves.

One can measure the reach of the phenomenon by observing that it depends neither on the trade nor on the nature of the tasks, but on a structure common to all expertise. Everywhere, the expert is the one who has ceased to do the elementary work because they have outgrown it; everywhere, the beginner is the one to whom this work falls, and who is thereby trained. This structure is so general that it runs through trades that have nothing else in common: an operating room, a cockpit, a firm, an engineering office obey the same principle of reproduction. It is precisely because this principle is universal that the automation of entry work poses, to all these professions at once, the same silent threat.

This mechanism was never designed. That is why it can be destroyed without anyone deciding to.

The quality paradox

A paradox makes the destruction harder still to perceive. In the short term, AI raises the average quality of beginners’ work. The assisted associate produces better research, the assisted resident makes fewer errors, the assisted young engineer delivers a cleaner calculation. Seen from immediate quality, automation is a net progress, and that is what all the indicators one looks at show. Nothing, in the dashboards, signals the problem; on the contrary, everything there seems to improve, which is the surest way not to see the danger coming.

For this rise in immediate quality is paid for by an erosion of deep training. The beginner whose work is done, or pre-done, by the machine obtains a better result without having traveled the path that alone deposits judgment. They deliver better and learn less. At the scale of a task, the gain is real; at the scale of a career, the loss is masked by the gain, which makes it almost impossible to perceive as long as one measures only the deliverable. One produces a generation that delivers, at equal age, higher-quality work than the previous one, and that will nonetheless have learned less to judge for itself.

The full paradox is then this: AI relies on the expertise accumulated by one generation, that of the experts trained the old way who know how to supervise and correct it, while removing the experiences that would let the next generation acquire the same expertise. It works because there are still experts to hold it; it compromises the manufacture of the experts who will have to hold it tomorrow. Every profession that automates its entry work thus lives on a capital of competence it no longer helps to renew, and this debt declares itself only when the generation trained the old way departs.

This delay between cause and effect is what makes the phenomenon so formidable. A profession can automate its entry work today and feel the consequences only in fifteen or twenty years, when the generation trained the old way retires without having been replaced in kind. Meanwhile, everything will seem to go for the best: higher-quality deliverables, reduced costs, apparently more capable beginners. The alarm will appear only when one must, for the first time, rely on experts who were not manufactured by the old mechanism, and one discovers they do not know what their elders knew. By then, the correction will demand a whole generation, for one does not recover in two years a training path that took fifteen.

AI lives on one generation’s expertise, while removing the experiences that would train the next.

What the path really deposited

To grasp the stake, one must be precise about what the path deposited, for it was not knowledge in the sense of information. Information, a beginner already has on leaving school. What they lack, and that only work gives, is of another order: the ability to recognize, in a concrete case, which of the learned principles applies and which misleads; the sense of danger, that inner alarm that makes an experienced practitioner know, before being able to explain it, that a situation is wrong. This sense is not taught, it is contracted, by repeated exposure to real cases and by the correction of one’s own errors.

Yet it is exactly this exposure that entry work provided, and that automation removes. Each failed then redone research, each first clumsy then corrected gesture, each first draft demolished by an elder was an occasion to contract this sense. The beginner learned not by succeeding, but by erring under supervision, in a frame where error was allowed because it would be caught. To remove the tasks is to remove the occasions to err usefully, and therefore the very mechanism by which judgment is formed. One loses not a mere exercise; one loses the right to the formative error, which is the heart of all expert learning.

This is what distinguishes deep training from a transfer of information. One can explain to a beginner why a given situation is dangerous; they will retain it, but they will not know it the way the one knows it who was once trapped. Knowledge transmitted by explanation stays external, available but cold; knowledge contracted by experience becomes a reflex, available under pressure, in urgency, when there is no longer time to reason. It is this second knowledge, the only one that holds in the fire of a real case, that the path deposited and that nothing, for now, replaces.

One learns to judge not by succeeding, but by erring under supervision. AI removes precisely those occasions.

Reproducing, not conserving

It would be easy, and wrong, to conclude that repetitive work must be artificially preserved to keep training beginners. This answer is defensive, hardly credible, untenable: no profession will give up a real gain to maintain a disguised pedagogical exercise, and no client, no patient, no passenger will agree to pay for inefficiency in the name of training. Regretting the old world trains no one. The real question is not how to retain what is leaving, but how to rebuild, differently, the function this work fulfilled without anyone ever having organized it.

And since this function was never designed, it can finally be designed. Where training was deposited by accident in entry work, it must now be carried by an explicit architecture of transmission, thought out, willed, inscribed in the very way automation is deployed. Make visible the reasoning that led to the result, and not only the result. Show the options discarded as much as the one retained, for it is in the sorting that judgment is formed. Organize a progressive delegation according to level. Transform verification into a learning exercise. Keep moments where the beginner formulates their position before seeing the machine’s, so they still think for themselves. The chance that trained experts has disappeared; it must be replaced by a device, and nothing forbids it to do better than chance.

There remains the question of the beginning, which was not rhetorical. A profession can perfectly well keep its knowledge and cease, at the same time, to produce the people capable of understanding it, contesting it and renewing it. It would even be the default outcome, the one toward which one slides if one decides nothing: professions that keep their methods and their archives, but where no one, in the end, really knows why they are right or how to make them evolve. Keeping a knowledge and reproducing those who carry it are two distinct things; automation excels at the first and threatens the second. The reproduction of a profession is not the conservation of its knowledge: it is the continuous manufacture of the minds capable of judging it, and it is that, not the tasks, that one must now decide to save.

Keeping a knowledge and reproducing those who can judge it are two distinct things. Automation succeeds at the first and threatens the second.