The Virtue of Letting Skills Die 

On August 12, 2026, Stanford’s Digital Economy Lab reported a striking divergence in the early-career labor market. Employment among workers ages 22 to 25 in highly AI-exposed occupations stood about 19 percent below where it would have been if it had kept pace with similarly aged workers in less exposed occupations. The researchers found no evidence of widespread economy-wide displacement and stressed that the pattern is descriptive rather than causal. Yet the divergence appears to operate primarily through reduced hiring. The deeper danger is that young workers may lose some of the work through which professional judgment used to be formed. 

A different body of evidence points in the opposite direction. In a field study of 5,179 customer-support agents, access to a generative AI assistant increased productivity by 14 percent on average and by 34 percent among novice and lower-skilled workers. The researchers found suggestive evidence that the system transmitted some of the practices of stronger workers to newer ones and helped novices move up the experience curve faster. AI can accelerate apprenticeship once a novice is inside an institution even as automation shrinks some entry-level tasks that once served as apprenticeship. A recent Trends in Cognitive Sciences review captures the tension: offloading cognition to AI can impede skill acquisition and contribute to skill decay, but the risks depend on how the technology is used. 

This paradox requires a more discriminating response than either nostalgia or enthusiasm. Some skills should disappear. A demanding practice does not become valuable merely because previous generations had to master it. Yet some inefficient-looking practices are formative: they cultivate mental models, habits of attention, error detection, and practical judgment. The right standard is therefore to preserve the function of a practice when that function forms the person, even when the technique itself can safely disappear. I call the capacity to make that distinction meta-competence: the prudence to know what to delegate, what one must still understand, when to distrust a fluent answer, and when to descend beneath a technological abstraction. Human flourishing in an AI-rich society will depend less on keeping every old skill alive than on knowing which losses make room for higher competence and which losses quietly destroy the path to it. 

Dead Difficulty and Formative Difficulty

The history of programming offers a useful model. Early programmers worked close to the machine, manually translating problems into long sequences of low-level instructions. When IBM released Fortran commercially in 1957, the abstraction was dramatic. IBM’s history records that a problem requiring as many as a thousand manually keyed program instructions could be reduced to forty-seven instructions in Fortran. Scientists and engineers could state problems in a form closer to mathematical notation while a compiler handled much of the translation. 

Few people would argue that programmers should have kept writing machine code simply because doing so was difficult. Fortran removed what we can call dead difficulty: effort that no longer cultivated a capacity necessary for the work humans were being asked to do. The abstraction freed attention for more valuable forms of reasoning and widened access to programming.  

Humans routinely reshape their environments to reduce mental work, a phenomenon psychologists describe as cognitive offloading. Writing, calculators, search engines, and software abstractions all move some cognitive burden outside the individual. Offloading can expand capability when it releases attention for more consequential judgment. 

Yet abstraction does not eliminate the need for understanding. A high-level programmer who knows nothing about what happens beneath the abstraction may struggle to diagnose a failure, reason about performance, or recognize when the tool’s behavior departs from the intended result. Research on skill retention shows why this matters: a quantitative review found substantial skill loss after periods of nonuse, with the amount depending on the type of skill and the interval without practice. Human-factors research on automation has likewise treated situation awareness, complacency, and skill degradation as central considerations in deciding what to automate and to what degree. 

That gives us the other category: formative difficulty. Formative difficulty produces something durable beyond the immediate output. It builds a mental model, a sense for exceptions, a capacity to explain why an answer is right, or the judgment to recognize when ordinary rules no longer apply. Education and professional formation have always mixed dead and formative difficulty. Students calculate by hand what software later calculates instantly. Junior employees produce drafts that supervisors rewrite. Trainees solve problems slowly that experts later handle with tools and shortcuts. Some of this effort should vanish as technology improves. Some looks inefficient precisely because learning is inefficient. The moral error lies in treating the exercise itself as sacred rather than asking what human capacity the exercise serves. 

AI Can Teach and Deprive 

The distinction becomes especially important because novice learning does not follow a simple rule that harder is always better. Research on the worked-example effect shows that novices can learn more efficiently from carefully structured examples than from unguided problem solving. A later experiment likewise found that example study outperformed problem solving alone for novices learning electrical-circuit troubleshooting. Other work shows that students who generate self-explanations while studying examples develop more example-independent understanding. At the same time, research on productive failure shows that struggling with complex problems before receiving full instruction can improve later transfer, and experiments on retrieval practice show that repeatedly retrieving knowledge can strengthen long-term retention more than additional study alone. The point is that the right kind of cognitive work matters for durable competence. 

Generative AI can participate in that formation when it is designed and used as a scaffold rather than a substitute for thought. The customer-support study is one example. Education provides another. In a randomized trial involving more than 700 tutors and 1,000 students, Stanford researchers found that a real-time AI system that modeled expert tutoring practices made students four percentage points more likely to master mathematics topics, with larger gains for students whose tutors had previously been rated lower. The system made expert patterns available at the point of need while keeping a human tutor engaged. 

AI can also conceal the absence of formation. A 2026 systematic review of generative AI in healthcare found recurring concerns about automation bias, deskilling, and effects on diagnostic reasoning. A novice analyst may likewise receive a polished AI draft and mistake fluency for sound reasoning, producing competent-looking work without the mental model needed to recognize a subtle error. A 2025 study of 319 knowledge workers found that higher confidence in generative AI was associated with less self-reported critical-thinking effort, while the nature of critical thinking shifted toward verification, integration, and stewardship. That shift can be healthy if verification and stewardship become real disciplines. It becomes dangerous when the worker has never developed enough independent competence to perform them well. 

Consider the first draft of an analytical memo. Writing the draft may force a novice to organize evidence, notice gaps, distinguish a claim from its support, and confront uncertainty. If AI writes it, the organization may receive a better document sooner and the novice a useful example of stronger work. Whether this counts as progress depends on what happens next. Does the analyst have to explain the reasoning in his own words, identify the strongest reason the draft could be wrong, trace important claims to evidence, and defend the recommendation against objections? Or does he simply forward polished prose that has never become knowledge? The same visible output can represent either accelerated formation or borrowed performance. 

Schools face the same choice. Requiring students to reproduce every old technique would confuse tradition with formation. Letting them bypass every demanding cognitive step would confuse performance with learning. A sensible AI policy therefore asks which experiences cultivate judgment and then redesigns those experiences around the new tool. A novice programmer might study an AI-generated solution, explain each step, predict what will happen when an assumption changes, and then solve a related problem without assistance. Research on scaffolded generative-AI coding support suggests that structured guidance can improve coding performance while supporting metacognitive self-regulation. The aim is to move difficulty to the place where it forms the capacities that still matter. 

Prudence About Delegation 

Meta-competence is therefore best understood as a form of prudence rather than a technical trick. The competent user of AI needs more than prompt-writing skill. He must know what understanding the task requires, what errors matter, what needs verification, and whether he could recover if the system failed. Older research on automation bias is relevant here. A systematic review found that people can over-rely on automated decision support and fail to detect new errors it introduces. A later review found automation bias even in single tasks when verification was complex. The risk is especially serious when the human role is reduced to passive approval, a version of the out-of-the-loop performance problem long studied in highly automated systems. 

Institutions can cultivate prudence deliberately. Before automating a routine task, a school or workplace should ask four questions. What does a novice learn by doing this task? Is that lesson still necessary for competent adult performance? If the task disappears, where will that lesson be learned instead? How will we test whether the person actually possesses the underlying competence rather than merely producing the right output with assistance? These questions make automation a decision about human formation as well as efficiency. 

Sometimes the answers will justify letting a skill die. Few adults need to perform every calculation by long division, and few programmers need to spend their careers translating formulas into machine instructions. Preserving those techniques universally would consume time better invested in more valuable capabilities. In other cases, the learning function must be rebuilt elsewhere. A junior professional who no longer writes every first draft might be required to diagnose an AI draft before revising it, reconstruct its argument from evidence without looking at the generated prose, handle the exceptions the system cannot resolve, or explain why an apparently plausible answer fails. These are not arbitrary hurdles. They are replacement exercises for capacities the old workflow once formed incidentally. 

Assessment must change as well. If a machine can reliably produce an ordinary answer, then human testing should increasingly emphasize abilities that make delegation safe: explanation, verification, transfer to a new case, recovery from error, and judgment under ambiguity. Meta-cognitive monitoring is trainable; a meta-analysis found that learning-strategy instruction can improve the accuracy with which learners monitor their own understanding. Meta-competence begins with an accurate sense of what one knows, what one does not know, and when assistance has exceeded one’s ability to evaluate it. 

The broader challenge in AI adoption at work is therefore institutional rather than merely individual. Organizations shape what people practice, what they are rewarded for noticing, and what kinds of mistakes they are expected to catch. A workplace that automates junior tasks without redesigning apprenticeship may enjoy short-term productivity while quietly consuming the human capital on which later expertise depends. A school that bans useful tools may preserve obsolete exercises while failing to teach students how to exercise judgment in the environment they will actually inhabit. Prudence requires a third path: remove dead difficulty, preserve formative difficulty, and create new forms of practice when the old ones disappear. 

Technological progress has always involved forgetting. We no longer expect educated people to retain every technique that once mediated competent work, and we should not treat the disappearance of a skill as a human defeat simply because the skill was hard to acquire. The real loss occurs when a society discards the exercise and fails to notice that the exercise was carrying a formative function.  

Human formation has never meant doing everything the hard way. It means becoming the kind of person who can recognize what the hard way was teaching. We should let many skills die. The virtue lies in doing so with enough prudence to know what must be reborn at a higher level. 

Image licensed via Adobe Stock.