When Performance Separates from Formation

This research note is based on the working paper "Unbundling Formation from Performance."

Artificial intelligence is usually discussed in the language of tasks. Can AI write the memo? Can it draft the email? Can it summarize the case? Can it solve the problem? Can it replace or augment a worker?

That is the right language for many labor-market questions. But it misses a second question that matters for institutions: what happens when the task was also the way people were formed?

Many human institutions did not maintain separate systems for producing output and forming people. They bundled the two together. A student's essay was not only a graded assignment. It was a practice through which reasoning, judgment, and intellectual responsibility were built. A junior analyst's first deck was not only a deliverable. It was a training surface where mistakes were corrected, standards were transmitted, and professional taste was formed. A legal memo was not only a piece of work. It was one of the ways a lawyer learned how to think like a lawyer.

In these cases, visible performance and hidden formation traveled through the same activity.

AI changes the economics of that bundle. It can often supply the visible performance without requiring the human process that used to carry formation. It can generate the draft, complete the assignment, produce the summary, clean the data, or write the analysis. The output remains. The formative path may not.

That is the central idea of my working paper Unbundling Formation from Performance.

The paper argues that AI's institutional impact should not be understood only as automation or augmentation. Its deeper effect is often unbundling: performance is separated from the human process that used to form capability, judgment, responsibility, or attachment.

This distinction matters because not all frictions are the same. Some friction is dead weight. It consumes time without forming anything. Some friction is deforming. It trains fear, passivity, hierarchy, or dependence. Those frictions should not be defended. But some friction is formative. It builds skill through repetition, judgment through correction, trust through presence, and responsibility through role-bearing.

AI is powerful partly because it can remove friction. The institutional challenge is that it may remove all three kinds at once. Dead friction disappears, which is good. Deforming friction may disappear, which is also good. But formative friction can disappear too, and its value is often harder to see at the moment of removal.

The immediate gain is visible. The delayed loss is not.

This creates a collective-action problem. Unbundling is usually cheap, private, and unilateral. A student, worker, firm, school, or platform can capture the short-term benefit of faster performance. Rebundling is harder. Rebuilding a new carrier for formation requires design, supervision, institutional commitment, and time. The benefit may appear years later, often to a different actor than the one who made the unbundling decision.

That asymmetry explains why the paper does not predict universal decline. The highest-risk cases are not simply the tasks most exposed to AI. They are the boundary cases where the old task still carried meaningful formative value, where AI makes unbundling easy, and where the actor deciding to unbundle does not fully internalize the future loss.

A legal firm that automates junior drafting may still produce good memos today. But if junior lawyers no longer pass through the disciplined process of issue spotting, revision, and partner correction, the firm may consume a stock of legal judgment it is no longer reproducing. A school that permits AI-written essays may still receive polished writing. But if students no longer struggle through argument, structure, and revision, the institution may preserve the artifact while weakening the capacity the artifact once helped form.

The same pattern can appear in workplaces, education, families, professions, and civic institutions. The visible output survives. The hidden formation pathway thins out.

The paper's model formalizes this intuition. It asks when an institution chooses to preserve a formative bundle and when it unbundles it. The key variables are the immediate gain from unbundling, the cost of maintaining the old formative carrier, the strength of the formative value, and the degree to which the deciding actor internalizes that delayed value.

The result is deliberately modest but important: the strongest bundles are not always the highest-risk bundles. Some are strong enough to defend themselves because actors can see and internalize their value. The most fragile cases lie near the preservation boundary. These are tasks valuable enough that losing them matters, but not visible enough, immediate enough, or privately captured enough to be protected.

This also reframes the design problem. The goal is not to preserve old tasks for their own sake. The goal is to ask what the task was carrying. If the task carried nothing but dead friction, remove it. If it carried deformation, dismantle it. But if it carried formation, the institution must decide whether and how to rebuild that function in a new form.

The core question is simple:

When AI removes the task, what happens to the person the task used to form?

That question is now becoming central across many domains. In education, it asks whether students still acquire judgment when answers are easy to obtain. In workplaces, it asks whether junior roles still form senior capability when entry-level work is automated. In professions, it asks whether judgment can be reproduced when routine drafting, diagnosis, or analysis is delegated. In families and communities, it asks whether convenience can replace service without weakening attachment.

The paper does not argue against AI. It argues for seeing the full institutional object. AI can improve performance and formation when used well. It can create new practice surfaces, better feedback, more individualized instruction, and less wasteful work. But those benefits do not appear automatically. They require rebundling.

The institutional challenge of the AI era is therefore not merely to ask what machines can do. It is to ask what human capacities our old activities quietly produced, and whether our new systems still produce them.

Performance can be automated faster than formation can be rebuilt.

That is the gap institutions now have to manage.