The apprenticeship paradox

Imagine a report arriving before the junior analyst has worked out what the question means. For her experienced manager, this can be a gift. He knows which assumption to challenge. One sentence catches his attention because he has seen that particular mistake survive three meetings and become an expensive decision. His colleague sees the same sentence. It looks like the others.

We often describe this arrangement by saying that AI will handle production and people will supply the judgment. But judgment has a history. Somewhere in the manager's history are the draft he had to defend, the discrepancy he almost missed and the colleague who asked a question he could not answer. His ability to evaluate the finished work developed through encounters with unfinished work.

This is the apprenticeship paradox. We may become better at using the expertise we already have while giving less attention to how its successors will acquire it. The immediate output improves. The question of how another person becomes able to understand, challenge and take responsibility for that output remains open.

The future of expertise depends on which experiences we preserve, improve and deliberately create as production becomes automated.

Learning research has long distinguished performance during practice from learning that survives and transfers. Soderstrom and Bjork's review shows why a smooth performance today can be a poor guide to what someone will remember or apply later. AI makes that distinction newly consequential. A polished result may now reveal very little about which decisions its apparent author understood.

The evidence does not support an inevitable story of decline. In Shen and Tamkin's small experiment, developers learning an unfamiliar Python library with AI assistance performed worse on an immediate knowledge test. The study could not establish what would happen over years of professional work. In a different setting, Brynjolfsson, Li and Raymond found productivity gains and evidence consistent with learning among customer-support workers using AI. The contrast makes the learning arrangement worth examining.

The old apprenticeship deserves examination too. Routine work was never automatically educational. Copying a report nobody explains, repeating an error nobody discusses or enduring humiliation called rigour can consume years without developing sound judgment. AI could remove that waste. It could expose a beginner to useful explanations and interesting problems much earlier.

What matters is the intellectual responsibility the learner gets to carry. Before seeing a recommendation, do they make a prediction? When comparing two plausible answers, can they explain what would make one appropriate? When a decision fails, do they encounter its consequences with someone who can help them understand what happened? Reviewing an answer can itself be demanding work. It requires enough knowledge to notice a problem and enough feedback to learn when one has been missed.

The cognitive scientists who described the assistance dilemma recognised this difficulty well before generative AI. Help can make learning possible, and help can displace an important part of learning. Withholding it can invite useful thought or leave someone hopelessly lost. There is no virtue in making a person struggle with an obstacle that teaches them nothing.

This changes the responsibility of the organisation. If every incentive rewards immediate completion, asking individuals to preserve their own development becomes a private task performed against the system. Time for explanation, safe responsibility and a second attempt has to exist somewhere in the design of work. An expert who quietly repairs every AI answer may protect delivery while leaving their colleagues unable to see what required repair.

Return to the report. Perhaps the manager asks the analyst which assumption she would test first. Perhaps she makes a prediction, examines the evidence and discovers why a convincing sentence can be wrong. AI has still saved time. The important difference is that some of the saved time has been invested in developing the person who will eventually have to make that decision without him.

Expertise will change as tools change. Remembering every procedure may matter less in some domains; recognising a consequential exception may matter more. We should allow that change rather than preserve every old task. But declaring judgment to be the human contribution does not create it. Somebody must take responsibility for the experiences through which another person becomes able to judge.