Why learn what the machine already knows?

When an AI tool can do a junior's work, the junior loses the reason to learn it. Most professions have taught their newcomers the same way. A first-year associate reviews documents. A junior analyst builds the spreadsheet a partner will glance at for thirty seconds. A new developer fixes small bugs in code nobody else wants to touch. The work is worth little. The learning is worth a great deal, and the low-value work is how the learning gets paid for. By the time the junior is trusted with hard problems, they have solved hundreds of easy ones.

Generative AI is very good at the easy ones. That is usually discussed as a question about jobs: will firms still hire juniors? This essay asks a different question, one that would matter even if every junior kept their job. If the tool already solves the problems at the bottom of the ladder, what is the reason to learn to solve them yourself?

What gets paid

In an earlier essay I described an AI tool by its frontier, the hardest problem it can solve for anyone who asks, and a worker by their ability, the hardest problem they can solve on their own. A worker whose ability lies below the frontier can still solve everything up to the frontier, by asking the tool.

Now follow that through to wages. Early on, when few firms use the tool, the employer who adopts it keeps most of the gain. Once the tool is standard, every employer can offer every worker the same floor, and competition passes it on to wages. At that point, the part of a worker's knowledge that the tool already has is worth nothing in the labour market: anyone can get it for the price of a subscription. Knowledge below the frontier stops being paid. Knowledge above it is paid more than before, because the tool's leverage multiplies it.

The wage–ability profile therefore develops a kink. It is flat from the bottom up to the frontier, and then rises more steeply than it used to. Workers just below the frontier are not paid less than before in absolute terms. The floor lifts them. What they lose is the price of what they spent years learning.

All or nothing

Consider a junior deciding whether to practise: to attempt a block of hard problems unaided rather than hand them to the tool. Practising costs something today, the output the tool would have produced on those problems. It pays off later, as higher ability.

Without AI, the decision is the same for everyone. Each rung climbed is paid, so if learning is worth its cost for one worker it is worth it for all. With AI, a rung below the frontier earns nothing. Practice pays only if it takes the junior past the frontier, into the region where ability is paid again. So the decision becomes all or nothing. A junior close to the frontier can jump the flat stretch and keeps practising. A junior further back cannot reach the paid region within the horizon, so for them practice is pure cost, and they stop.

With the illustrative numbers in Figure 1, every worker practises without AI. With a tool whose frontier covers 60% of the job's problems, practice stops for every junior who starts below 0.45 on the same scale. They are more productive than their predecessors on day one, because the tool answers for them, and they will be exactly as capable in five years as they are today.

Every model release raises the bar

The threshold for practice moves one for one with the frontier. Each time the tool learns to solve a harder class of problems, the flat stretch lengthens by the same amount, and the band of juniors for whom learning no longer pays widens with it.

This is the human-capital side of the valley. In the earlier essay, the workers who gained least from AI were those whose knowledge sat at its frontier. Here, the workers with the least reason to learn are those who sit some distance below it, and the distance grows as models improve.

Crutch or tutor

Two forces push the other way. The first is leverage: because the tool multiplies the output of those above the frontier, the prize for getting there is larger, and more juniors find the jump worth attempting. The second, and more important, is how the tool is used. A tool that hands over answers is a crutch. A tool that explains its answers, asks the junior to try first, or shows where their attempt went wrong is a tutor, and makes each block of practice teach more. In Figure 1, a tutoring mode that raises learning by half lowers the threshold from 0.45 to 0.35. It narrows the band of juniors who stop learning. It does not close it.

Whether AI acts mainly as a crutch or mainly as a tutor in a given occupation is therefore a race: between how fast it speeds up learning and how fast the frontier moves. Nothing guarantees which side wins.

Who pays for the rungs?

So far the junior has chosen. In most firms, the employer decides how work is allocated, and here a sixty-year-old argument from Gary Becker bites. Skills that are useful at any employer are general human capital. A firm in a competitive labour market cannot recover the cost of teaching them, because once the worker is skilled, another firm can hire them away at the higher wage. So the firm has no reason to pay for general training, and the worker must pay for it, typically by accepting a low wage while learning.

Entry-level jobs have long solved this by bundling. The junior does low-value work at a low wage, and the training comes with it. AI breaks the bundle. The low-value work can now be done by the tool, at a fraction of the cost, and letting a junior struggle with it means giving up output the tool would have produced for free. A firm that assigns the easy problems to the tool loses nothing it can measure this year. What it loses, the expert that junior would have become, accrues mostly to whoever employs them later. Frictions in the labour market, which let firms keep part of the return to training, soften this, as Daron Acemoglu and Jörn-Steffen Pischke showed. They do not reverse it.

The cost appears later and elsewhere. The workers who end up above the frontier are the ones whose judgment AI cannot replace: they catch the plausible wrong answer, handle the problem nobody has seen, and supervise the tool. Research Note 002 shows that this judgment is precisely the skill AI rewards most. If fewer juniors climb past the frontier, the stock of such people shrinks, just as demand for them rises.

What follows

For firms, the implication is that training is now a decision rather than a by-product. It used to come free with the junior's work. Now it has to be chosen: time set aside for unaided practice, tools configured to explain rather than answer, and some acceptance that a junior working without the tool is slower this quarter. Firms that skip it will find, in a few years, that they can buy the tool but not the people who know when it is wrong.

For juniors, the advice is to aim at what lies beyond the frontier rather than at what lies just below it. Mastering problems the tool already solves is not worthless, because it is the route to judgment. But it pays only as a means of getting past the frontier, not as an end.

For research, the model makes a sharp prediction. Among juniors whose tasks lie below the frontier, practice and skill growth should fall when the frontier rises, and the fall should be smaller where the tool is used as a tutor. Randomising a tutoring mode, in which the tool explains instead of answering, would measure the size of that offset directly. If AI-assisted juniors whose work lies entirely below the frontier learn faster than their predecessors without any tutoring, the argument is wrong.