Improvisation as Expertise

In the AI Story

The claim extends Suchman's situated action framework from her original photocopier studies to the broader question of what expertise consists of. Julian Orr's research on Xerox field technicians documented it directly: the most effective repair workers were not the most procedure-faithful but the most responsive to specific circumstances. They heard jams forming before error lights appeared. They felt through the vibration of the machine's frame whether paper was feeding correctly. Their knowledge was experiential, accumulated through thousands of encounters with specific machines, and it resisted formalization because it was constitutively bound to the circumstances that produced it.

The pattern generalizes. The surgeon who encounters unexpected adhesions adapts her technique, drawing on embodied knowledge and accumulated experience of what works in similar but never identical circumstances. The air traffic controller who sees aircraft converging on an unexpected vector reconfigures the traffic pattern in real time. The jazz musician's phrasing responds to the room's acoustics, the pianist's modulation, her own fatigue. In every case, the practitioner's intelligence manifests in the departure from the plan — in the moment when the situation demands something the plan did not specify, and the practitioner produces a response that addresses the demand.

The improvisation is not random. It is informed by everything the practitioner has encountered before, calibrated by feedback from the current situation, and directed by a form of judgment that operates faster and more reliably than deliberate analysis because it has been shaped by accumulated experience. It is, in Donald Schon's phrase, knowing-in-action — competence that cannot be separated from the performance through which it is expressed.

The implications for AI are specific. When AI handles implementation tasks that look routine but contain embedded improvisational moments — the ten minutes of unexpected behavior in the four-hour block of dependency management — it is not just eliminating tedium. It is eliminating the occasions on which improvisational expertise develops. The practitioner who receives AI-generated solutions without engaging in the improvisational work that produced them possesses the outputs without the capacity; she can deploy what the machine produces but cannot reliably recognize when the next output will fail, because the recognition capacity is built through the exact kind of engagement the machine has now absorbed.

Origin

The claim emerges from the convergence of multiple research traditions: Suchman's situated action (1987), Orr's ethnography of technicians (1996), Schon's work on reflective practice (1983), Klein's naturalistic decision making (1980s-1990s), and the broader literature on expertise developed by Ericsson, Dreyfus, and others.

The application to AI specifically emerges in the On AI cycle as a way of naming what is lost when AI automates practices that look routine from the outside but contain improvisational moments whose formative function is invisible in output metrics.

Key Ideas

Expertise is not rule-following. The most competent practitioners are the most improvisationally responsive, not the most procedure-faithful.

Disciplined improvisation. The improvisation that constitutes expertise requires years of practice to be possible; it is structure adapting to contingency, not chaos.

The domain-specific repertoire. Each domain produces its own specific repertoire of responses; the expertise is bound to the domain in which it developed.

Invisible from outside. Improvisational expertise looks like rule-following to observers who cannot see the accumulated experience beneath the performance.

AI eliminates the developmental occasions. Automating implementation tasks eliminates the improvisational moments through which expertise develops, producing practitioners who possess outputs without capacities.