When AI Does the Work, Who Gets the Practice?

The hidden cost of a correct answer

A software developer reviewing code can either accept an AI assistant's answer or work through the problem independently. The assistant may be faster and more accurate. Yet the independent attempt may help the developer learn to spot the next bug without help.

That tension is the subject of Today's Work, Tomorrow's Workforce: Incentivizing Learning in AI-Assisted Work, an arXiv preprint by Moritz Diener, Daniel Hendriks and colleagues at Karlsruhe Institute of Technology and IBM Research. The paper presents a mathematical model, not a workplace experiment. Its examples illustrate what could happen under specified assumptions; they do not measure how often workers lose skills or establish the price of training.

The authors build on earlier models of the choice between accepting AI advice and solving a task oneself. They add a sequence of tasks and a changing level of human skill, represented by unaided accuracy within one family of tasks. In their model, independent solving takes effort and can raise skill, even when the answer is wrong. Accepting the AI's answer saves effort but can allow skill to decay. Neither change is guaranteed on any particular task.

Why looking ahead may not be enough

The paper compares a worker who considers only the immediate task with one who also values future returns from skill. For the first worker, solving independently is worthwhile only when the accuracy advantage over AI compensates for the effort. A forward-looking worker may instead practice even when the AI is more accurate now, because improved skill can pay off on later tasks.

But foresight does not always make practice worthwhile. A novice may be so far below the AI's accuracy that the immediate costs of practicing outweigh the expected future gains. In the model, repeatedly accepting AI answers then prevents the novice from reaching the level at which independent work would pay for itself.

Experts face a different risk. If their accuracy advantage over AI is too small to justify the effort of solving, they can rationally accept the AI's answers even when they expect many more tasks. Their own skill can decline while the accuracy of submitted answers stays at the AI's level. The authors call these outcomes a low-skill trap and expert deskilling. They are consequences of the model's rules, not observations of the paper's participants: there were no human participants.

The distinction matters for evaluation. A measure that checks only whether the more accurate party answered the current task can rate a novice's reliance on AI as appropriate, while missing the practice not taken. Conversely, it can label independent practice as underreliance even when that practice has future value. The authors suggest tracking unaided accuracy alongside AI-assisted output over time.

Paying for practice

The paper next considers an employer that pays a bonus for correct answers produced independently. An earlier effort-based incentive covers the expected cost of solving. In the model, that can persuade a worker who is more accurate than the AI to work independently, but it does not necessarily help a less accurate novice: the novice still gives up the reward associated with the AI's more accurate answers.

The proposed learning incentive covers both the effort and, when the worker is less accurate, the expected reward lost by not accepting the AI's answer. It is calculated for a worker focused on immediate payoff and assumes ties favor independent work. The actual payment is made only for a correct independent answer, although an incorrect attempt can also teach under the model.

Support need not always continue forever. If practice takes a worker to a level where solving independently pays without a bonus, later practice can sustain itself. If it does not, the worker may return to AI reliance and need recurring support. In one illustrative scenario, keeping an expert at a higher skill target costs less in long-run incentive payments than allowing skill to fall and repeatedly paying to rebuild it. That comparison depends on the assumed learning and decay rates; a higher target is not always cheaper.

What the model leaves out

The results depend on a deliberately narrow picture of work. The model assumes skill grows only through independent solving, never through engaged use of AI. Other cited research suggests assistance design can matter: guided tutoring or explanations may support learning, while unrestricted help can harm it. The model also assumes the employer can observe skill and verify that an answer was produced without AI, and that workers know their own accuracy. Those assumptions are consequential for feasibility and privacy.

Incentive spending is not the whole social cost. Independent practice by a less accurate worker can produce more errors than using AI; the people affected by those errors may not be the people who later benefit from the worker's expertise. The authors also note that skill assessments could be used for discipline and argue that workers should have a say in setting targets and using assessments.

The preprint calls for longitudinal research to estimate learning and decay, test whether incentives work, and assess how practice affects both workers and outcomes. Its practical challenge is not to abandon AI assistance, but to ask whether a system that delivers good answers today is also preserving the human capabilities needed tomorrow.