Decoupling of Learning and Producing
Eleanor Gibson's research on perceptual learning established that the organism learns through the same engagements through which it acts. The infant learning to crawl does not first crawl in a simulation and then crawl for real; she learns by crawling. The builder who debugs a race condition does not first study race conditions abstractly and then encounter them; her perceptual differentiation develops through the debugging itself, with the productive activity and the developmental activity fused. The apprenticeship model across every skilled trade in human history embedded learning inside production: the apprentice learned by doing the work, and the friction of the work was the mechanism of the learning. The AI-augmented environment dissolves this coupling. The builder can now produce without undergoing the exploratory engagement that historically drove differentiation. The production looks identical — the output is shipped, the feature works, the metric registers. The learning has not occurred. Whether this matters depends on whether the developmental function of the old friction can be replicated through deliberate means, or whether a generation of builders will reach senior roles with surface competence that rests on perceptual foundations that were never fully built.
In the AI Story
The coupling of learning and producing in skilled trades was not a pedagogical design — it was the ecological default. Apprentices did real work because real work was what needed doing, and the tools available made the work cheap enough to delegate but expensive enough that the delegation required actual engagement. The apprentice who watched without doing did not become a journeyman; the apprentice who did, did.
The pre-AI software environment preserved this coupling. The junior developer who debugged production issues developed, through the debugging, the perceptual sensitivities that distinguished senior from junior performance. The senior developer who mentored juniors transmitted not stored knowledge but attentional guidance — pointing the junior toward invariants the senior had learned to notice, so the junior's own engagement with the next debugging episode would be more perceptually productive.
AI-augmented production removes the activity through which learning historically occurred. The builder describes the desired outcome; the AI produces the implementation; the builder reviews the output. Review is a perceptual activity, and it develops certain sensitivities — but they are the sensitivities of reviewing AI output, not the sensitivities of direct engagement with the system's affordance structure. The perceptual skills of reviewing and the perceptual skills of implementing are both real, but they are different skills, and the question of whether one can substitute developmentally for the other is empirically open.
Gibson's framework is pessimistic about the substitution. Perceptual differentiation, on his account, depends on the organism's active exploration of the environment's structure. The review of another agent's exploration — whether the agent is a senior developer or an AI — does not generate the transformational samples from which invariants are extracted. The reviewer gains information; the reviewer does not necessarily gain differentiation. Whether structured practices can be designed to reintroduce the coupling deliberately — whether, for instance, AI-augmented teams can include protected practice periods during which junior builders engage directly with systems without AI mediation — is the practical pedagogical question that the next decade of engineering education and apprenticeship design will have to answer.
Origin
The diagnosis emerges from this book's application of Eleanor Gibson's perceptual learning framework to the AI-augmented builder's environment, read through Edo Segal's account of the December 2025 threshold and its consequences for the developmental trajectory of software engineering.
Key Ideas
Historical coupling. Across every skilled trade, learning and producing were coupled through the same activities; the coupling was the ecological default, not a pedagogical choice.
AI's dissolution. AI-augmented production can proceed without the exploratory engagement that historically drove learning.
Reviewing is not implementing. The perceptual skills developed by reviewing AI output differ from those developed by direct engagement with the system's affordance structure.
The substitution question. Whether deliberate structures can replicate the developmental function of coupled engagement, or whether surface competence will increasingly rest on unbuilt foundations, is an open empirical question.
The apprenticeship design problem. Engineering organizations face a novel pedagogical task: building environments that afford both production (which AI accelerates) and learning (which AI bypasses).
Debates & Critiques
The practical debate divides those who believe deliberate practice structures can substitute for the coupling AI dissolves from those who believe the coupling was itself the mechanism and cannot be replaced by designed exercises. The empirical answer will depend on cohort studies of builders who entered the profession after December 2025 and will not be fully available until those cohorts reach senior technical roles — likely in the early-to-mid 2030s. The wager of this book is Gibsonian: authenticity of engagement matters, simulated friction does not substitute, and organizations that do not deliberately preserve occasions for direct engagement will discover the cost when the smooth environment demands rough perception.
Appears in the Orange Pill Cycle
Further reading
- Eleanor J. Gibson, Principles of Perceptual Learning and Development (1969)
- Harry Collins, Tacit and Explicit Knowledge (2010)
- Hubert Dreyfus and Stuart Dreyfus, Mind Over Machine (1986)
- Matthew B. Crawford, Shop Class as Soulcraft (2009)
- K. Anders Ericsson, Peak (2016)