The Unbundling of Expertise

In the AI Story

The bundling was never logical, only contingent. Ryle's dispositional analysis identifies mental capacities as clusters of tendencies, and there is no a priori reason why the disposition to produce must travel with the disposition to evaluate. The two were bundled historically because human brains tend to develop them together through training — the programmer who learned to write code also learned, through the same process of iteration and correction, to judge what code was worth writing. The bundling was an artifact of the developmental process, not a necessary feature of the cognitive architecture.

Claude's dispositional profile reveals the contingency. Training on a corpus of code produces extremely reliable productive dispositions — code that compiles, runs, handles common cases — without producing the evaluative dispositions that distinguish good code from bad, appropriate solutions from merely plausible ones. The same dissociation appears across domains: Claude can produce fluent prose without reliably judging whether the prose is true (see the Deleuze error), can generate arguments without evaluating whether they are sound, can draw connections without assessing whether the connections hold.

The practical consequence is a division of labor structured by dispositional profiles. The machine handles production; the human handles evaluation. The senior engineer in Trivandrum discovered this directly: the twenty percent of his work Claude could not handle — the judgment about what to build, the architectural intuition about what would break, the taste separating useful features from merely plausible ones — was the part that mattered. The knowing how of judgment is what remains irreducibly human, not because of metaphysical privilege but because the developmental history required to build it is not one AI training approximates.

The educational consequences follow. A system that cultivates productive dispositions without evaluative ones now produces graduates with the specific competence machines replicate. A system that deliberately cultivates evaluative dispositions — judgment, questioning, the discrimination between plausible and true — produces graduates whose skills retain value. The teacher who grades questions rather than essays has grasped the unbundling pedagogically. The institutions that have not made this shift are training students for a competence the market no longer rewards.

Origin

The term 'unbundling' is common in platform economics, but the dispositional framing is introduced in the Ryle volume's chapter 3 to describe what AI has made visible in the structure of cognitive work. The underlying analysis draws on Ryle's treatment of knowing how as a cluster of distinct dispositions that can be exercised or absent independently.

Key Ideas

Contingent, not necessary bundling. Productive and evaluative dispositions traveled together historically because of how humans develop them. The machine shows the bundling was not logically required.

Reliability profiles diverge. Claude's productive dispositions are robust; its evaluative dispositions are weak. The same system possesses one cluster fully and the other partially.

Economic inversion. The rarity of production collapses; the previously invisible rarity of evaluation becomes the scarce resource the market rewards.

Educational implication. Systems that test production alone now test what machines do better. Systems that develop evaluation produce the skills that remain valuable.

Debates & Critiques

The question is whether the unbundling is temporary — an artifact of current AI architectures that will be resolved as systems improve — or structural, a permanent feature of the cognitive landscape. The Ryle volume takes no definite position, but notes that the self-correction disposition (essential to evaluation) requires precisely the feedback-and-correction developmental history that AI training does not provide, and that this structural gap may persist across architectural generations.