The Missing Practice: Productive Failure and Verification

When routine work disappears, the loss is not measured in lines of code alone. The focus here is productive failure—the struggle and revision that turn an imperfect attempt into transferable competence—and the related cost of verification. This distinction helps us judge whether an AI-assisted result belongs to the learner or merely passed through them.

The difference matters because the easiest way to complete a task is not always the best way to learn how to do it. Generative AI can make a plausible answer appear before a beginner has formed a mental model of the problem. That can improve short-term performance while weakening the ability to recognize an error, explain a choice, or work without the tool. The danger is not that every use of AI prevents learning. It is that visible success can conceal the disappearance of the practice that makes later judgment possible.

Why easy success can teach less

A beginner does not become competent by collecting correct answers. They become competent by forming explanations, testing them against reality, noticing where they fail, and revising them. A low-stakes task is valuable when it creates that loop. The learner has enough room to make an error, enough feedback to understand it, and enough connection to real work to see why the correction matters.

The software-engineering interviews describe what happens when that loop becomes shorter or disappears. All eight junior participants reported that, compared with the period before widespread GenAI use, they could achieve good results more easily in coursework or preparation. The easier result was not always experienced as progress. One participant said, “I don’t know what not to do. I’ve only seen the good examples.” Another summarized the deeper problem as, “I don’t know what I don’t know.”

Those statements describe a missing reference point. Someone who has only seen polished solutions may know how to reproduce a pattern without knowing which assumptions made it work. They may recognize the surface form of a correct answer but not the conditions under which it fails. In software work, that difference can remain invisible until the requirements change, the data is unusual, or the system behaves in a way the generated answer did not anticipate.

The study interprets this pattern through two related ideas from learning research. Productive failure describes the possibility that struggling with an unfamiliar problem before receiving a canonical solution can produce deeper conceptual understanding than receiving the solution first. Desirable difficulties describes conditions that make immediate performance harder but can improve later retention and transfer. These concepts do not say that all frustration is educational or that a learner should be left without help. They identify a useful tension: a process that is inefficient for today’s answer may be valuable for tomorrow’s judgment.

That tension is easy to miss in a workplace because production and development share the same task.