Borrowed expertise: Why AI’s productivity boom may not survive the generation that built it

The productivity numbers from generative artificial intelligence (AI) are remarkable, and they are real. Across knowledge industries, output per worker is rising in ways that would have been hard to imagine three years ago. The natural conclusion, drawn by many, is that we are at the early stage of a long economic boom, with the gains compounding as adoption deepens.

But this conclusion rests on a hidden assumption, and that assumption is becoming less true every day. The productivity miracle we are observing is produced, in large part, by people whose expertise was built before AI existed. They are extraordinarily good at directing these tools because they have spent careers training the kind of judgment that knows what to ask, what a good answer looks like, and where a confident-sounding model is likely to be wrong. They are extracting value from AI in part because they paid the developmental cost that AI now lets others avoid.

The conditions that produced this expert class are eroding. The pipeline that would replace them is being attenuated. And the productivity story we tell ourselves about the next 20 years assumes the steady arrival of new senior experts who will, in fact, not arrive on schedule.

The familiar AI and inequality concerns are about access and displacement: who can use the tools and whose jobs are automated. The argument here is different. It is about cognitive development, the conditions under which expertise gets built, and what happens when those conditions are quietly removed for an entire generation. The corollary is what such a society does about innovation, paradigm change, and the creation of new knowledge, which I will explore in a second piece.

The empirical literature on AI’s productivity effects points consistently in one direction. In a study of 5,000 customer service agents at a Fortune 500 software firm, Erik Brynjolfsson, Danielle Li, and Lindsey Raymond found that AI assistance produced substantial productivity gains. This included a 34% improvement for novice and low-skilled workers but minimal effects for the most experienced. Fabrizio Dell’Acqua and colleagues found similar patterns in a field experiment with consultants at Boston Consulting Group (BCG): AI improved both speed and quality of output, again with the lowest-performing consultants showing the biggest gains. Read carelessly, these results look like an argument against the developmental concern.

They are not. To understand why, it helps to be precise about what generative AI is doing in these settings.

In customer service, in much of consulting, and in routine legal, medical, and accounting work, the task is largely information retrieval. The right answer, or something close to it, exists somewhere in the model’s training data or in the firm’s internal documentation, and the difficulty lies in finding it, packaging it appropriately, and applying it to the case at hand. Generative AI is, predictably, very good at this kind of work. The Brynjolfsson, Li, and Raymond study itself describes the mechanism as one in which the AI captures and disseminates the tacit knowledge of the firm’s most productive agents, allowing newer workers to move down the experience curve faster. This is information retrieval, broadly construed: The model has learned what the experts know and can serve it back to those who do not.

This kind of work is real, valuable, and a substantial share of the modern knowledge economy. It is not, however, the work that defines expertise in the strict sense. The work of a research scientist generating a new hypothesis, of an engineer designing a technology that does not yet exist, of a physician facing an atypical presentation, or of a lawyer arguing a case the courts have never seen, is not principally information retrieval. It is decisionmaking under genuine uncertainty. It requires creativity in how the problem is framed, the integration of knowledge across disciplines, the application of life experience and what Michael Polanyi called tacit knowledge, the kind of knowing-more-than-we-can-tell that is acquired through immersion rather than instruction. It involves the generation of new knowledge rather than the recall of old. On this kind of work, AI is far less helpful, and the user’s domain expertise is far more decisive in determining whether AI helps at all.

The second study captures a piece of this directly. The same paper that documented impressive productivity gains within the AI’s capability frontier also documented a sharp deterioration in performance, of about 19 percentage points, when consultants used AI on a task that fell outside that frontier and that the AI could not solve correctly. The model produced confident-sounding answers. The consultants, lacking sufficient domain expertise to detect the errors, followed them—same tool, different task, opposite effect. The authors named this the “jagged technological frontier,” and the diagnostic point is precisely that the user often cannot tell which side of the frontier any given task falls on. Telling requires the very expertise that the productivity studies, by their design, mostly bracket out.

The result is a divide that the productivity studies, focused on the routine end of the spectrum, are not designed to detect. AI compresses the gap at the bottom of the skill distribution, where the work is dominated by information retrieval, and widens it at the top, where the work is dominated by judgment under uncertainty and the creation of new knowledge. The economy may look more equal in the short run while becoming more dependent than ever on a thin layer at the top whose replacement is no longer being produced.

The expertise of today’s senior knowledge workers was built through what cognitive scientists since K. Anders Ericsson have called deliberate practice: effortful engagement with hard problems, repeated exposure to honest feedback, and the slow, mostly invisible accumulation of pattern recognition that Daniel Kahneman associated with expert intuition. The lawyer who can spot the issue in a contract before reading it carefully, the radiologist who feels something is wrong with an image before locating the lesion, and the economist who senses that a clean-looking model is missing a variable all developed that ability the same way. None of it came from being told the answer; it came from working through cases unaided, getting things wrong in instructive ways, and being corrected.

The hard part is that developmental work is invisible in the work product. A junior analyst who writes a memo with substantial AI assistance and a junior analyst who writes the same memo unaided turn in indistinguishable artifacts. From the perspective of the firm, the supervisor, and even the analyst, the second mode of work looks like a needless slog. The analyst is “more productive” with AI in any measurement that takes the memo as the unit of output. What the measurement misses is that the unaided slog is what builds the analyst into someone who can, 10 years later, smell a flawed argument in a partner’s draft. The market does not price this developmental externality. It is invisible until it isn’t, and by the time it isn’t, it is too late.

Commercial large language models make this problem worse, not because of any malice in their design, but because of what their incentives select for. The product is helpfulness. The user-satisfaction metric rewards giving answers, not refusing to answer in the service of the user’s long-term cognitive development. There is currently no commercial model whose product is productive struggle. There may someday be one, but it is not the world we currently inhabit.

The most serious counterargument is that AI can be the tutor the world has never had. Benjamin Bloom’s classic finding, that one-on-one tutoring outperforms group instruction by roughly two standard deviations, has stood for 40 years as an unattainable benchmark. The pitch, made forcefully by Andrej Karpathy and the team at Khan Academy, among others, is that AI can deliver Bloom’s two-sigma at scale. AI does not have to skip the developmental work; it can scaffold the developmental work more effectively than any human teacher could.

In principle, this is correct, and in some specific deployments it is already happening. The question is whether the deployed systems most students and workers actually use have this character. As a university professor, I have a relatively direct view of how students actually use these tools, and what I see is not the patient Socratic tutor of the optimistic vision. The systems they reach for are answer engines. The student with a problem set asks for the solution, not for a hint. The student writing a paper asks for paragraphs, not for feedback on a draft. The pattern is consistent across courses and across cohorts. The systems are doing what they are designed to do, which is to give the user what the user has asked for, and what the user has asked for is the answer. A model that responded to a homework question with “I will not give you the answer; here is a question that should help you find it yourself” would lose that user to a competitor on the first turn.

The pedagogical AI tutor is a possible product. It is not, in any serious volume, the actual product. The developmental costs are being incurred in the gap between what the technology could do in principle and what the deployed systems actually do. Closing that gap is both a design problem and a policy problem. It is not solved by pointing out that it is solvable.

The expert class that is currently extracting value from AI will age out. Some will retire, some will slow down, and all will eventually leave the workforce. The pipeline that produced them is, in real time, being attenuated.

In 20 years, the cohort that should have been the senior expert class of the 2040s will instead be a cohort that grew up with AI doing the cognitive work that builds expertise. They will be functional; they will be productive on familiar tasks, perhaps highly so; but they will be substantially less able to handle the tasks where expertise actually matters: the novel, the ambiguous, and the unprecedented. Their judgment will be thinner, because the conditions that build judgment were not present in their formation.

To be clear, this is not a story about lazy younger workers. The cohort I am describing did not choose its developmental environment. We did, by deploying these tools without regard to their effect on cognitive formation, and by allowing the market to shape the deployment.

The implication for the productivity story is uncomfortable. The current gains assume a population of expert prompt-writers and judgment-providers who can extract value from these tools. That population is being drawn down without replacement. The productivity boom is, in this sense, partially a transfer from the future to the present. We are spending an inheritance of expertise that took decades to accumulate, and the rate at which we replenish it is no longer matched to the rate at which we are spending it.

The same pattern is now visible inside firms, and the firm-level version may be the clearest empirical evidence available. Recent studies suggest that companies are responding to generative AI by reducing their hiring of junior staff while continuing to grow their senior ranks. Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen, using payroll records from millions of U.S. workers, found a roughly 16% relative employment decline for workers aged 22–25 in the most AI-exposed occupations since the widespread release of generative AI tools, while employment for older workers in the same occupations remained stable. Seyed Mahdi Hosseini and Guy Lichtinger, analyzing resume and job-posting data covering 65 million workers across 280,000 firms, found that companies actively adopting generative AI sharply reduced junior hiring relative to non-adopting firms, while senior employment in the same firms remained mostly unchanged. They describe the pattern as a “seniority-biased technological change,” and the junior decline is driven not by layoffs but by firms simply ceasing to hire at the entry level.

In the short run, this is rational at the level of the individual firm. The work that juniors used to do—the routine drafting and analysis and information retrieval—is precisely the kind of work AI now performs at a fraction of the cost. A firm that does not need as many juniors will, naturally, hire fewer of them.

In the long run, this is a problem the firm has not internalized. The juniors a firm does not hire today are the seniors that firm will not have in 15 years. Senior expertise inside an organization is not only a stock of knowledge; it is the residue of years of junior work performed inside that specific organization, the slow accumulation of judgment about how the firm in particular handles its clients, its problems, and its institutional history. That residue cannot be acquired from outside. By collapsing entry-level hiring now, firms are quietly dismantling the mechanism by which they have historically produced their own future leadership.

The collective version is worse than the individual version. If every firm in an industry independently decides not to hire juniors, the entire industry’s senior pipeline empties out together, and there is nowhere for any firm to recruit the senior expertise it will need, because the cohort that would have produced it was never trained. This is a coordination failure, and like other coordination failures, it cannot be solved by any single firm acting alone. It is a problem for industry associations, for credentialing bodies, and for policy.

There is a story from the history of computing that bears directly on this. Early programmers worked under brutal hardware constraints. Memory was expensive, processor cycles were scarce, and writing efficient code was a form of intellectual craft. Programmers thought hard about algorithms because the alternative was a program that did not run. The result was several decades of remarkable innovation in algorithm design, much of it driven by the constraint itself.

Then, hardware became cheap, faster than software got demanding. The pressure to write efficient code largely disappeared from mainstream commercial programming. The craft survived in niches: graphics, embedded systems, scientific computing, and high-frequency trading. The median programmer no longer needs to think the way Donald Knuth thought and largely does not. Code is fine. Programs run. The hardware absorbs the inefficiency. The cognitive innovation that was forced into existence by the constraint is no longer forced, and it is no longer produced at anything like the historical rate.

The analogy to thinking under inference-cheap conditions is necessary to keep in mind. Hard problems used to force innovation, because the alternative to innovation was failure. The cost of taking the long way, of not finding the elegant solution, of brute-forcing through, was that you did not ship. Now the cost is much lower. AI will brute-force its way through problems that would previously have required a new idea. The work product is acceptable. The work is acceptable. But the new ideas do not get generated at the historical rate, because the conditions that generated them are gone.

This is not an argument that AI prevents innovation. It is an argument that AI changes the marginal incentive to innovate cognitively, and that the long-run aggregate effect on the rate of genuine novelty is plausibly negative. The visible productivity rises. The invisible production of new ideas slows. The two effects are not contradictory, and the second is much harder to measure than the first.

There is a sharper version of this concern visible in current programming practice itself. Generative AI is now extraordinarily good at writing code in established languages like Python. The reason is straightforward: It has been trained on an enormous corpus of human-written Python, much of it produced by skilled programmers over decades. For routine snippets, the model is at least as good as most human programmers and considerably faster. Now suppose a new programming language is introduced tomorrow, one that has no significant body of code yet written in it. AI would be largely useless in this language until a corpus accumulates. The work of writing, debugging, and idiomatically applying the new language would fall to humans, and the humans best positioned to do it would be those who learned to code the hard way. They have the underlying understanding of what programs are, what abstractions matter, and what tradeoffs are involved in language design. They can recognize what Python does well, what it does poorly, and where the new language might genuinely improve what came before. The programmer whose Python skills consist largely of prompting an AI does not have those underlying capacities. They are excellent at routine tasks within an established corpus and helpless outside of them. They cannot evaluate whether a new language is worth adopting, because they do not understand the language they already use well enough to compare.

The pattern is not specific to programming. Generative AI extends human capability most powerfully in domains with mature corpora. The leading edge of any field, the place where new tools and new framings are introduced, is precisely where the corpora does not yet exist. The people who can work at that edge are the people who have built the underlying capacity that does not depend on a corpus. We are reducing the rate at which we produce them.

The deeper question, which the productivity framing tends to miss, is what happens when AI knows essentially everything we already know, and we mostly use it to recombine and repackage what we already know. Where does new knowledge come from?

The optimistic answer is that AI does the legwork, and humans do the actual thinking at the frontier. As a description of how the work would ideally divide, this is coherent. As a description of what is actually happening, it requires a population of humans capable of doing the thinking at the frontier. But, again, we are reducing the rate at which such humans are produced.

There is now suggestive empirical evidence that AI assistance, even where it appears to enhance creativity, simultaneously reduces the diversity of the resulting ideas. In a controlled experiment published in Science Advances, Anil Doshi and Oliver Hauser found that writers given access to generative AI produced stories that were rated more creative on an individual basis but were measurably more alike in aggregate, with collective diversity falling by roughly 10%. The Dell’Acqua and colleagues paper documented an analogous compression of idea diversity among the BCG consultants. The finding is consistent across settings: AI lifts the average and narrows the variance. Innovation, however, comes disproportionately from the tail. A productivity gain that is purchased by a reduction in the variance of ideas is not a free lunch for the rate of discovery; it is a trade-off, and the side of the trade-off being paid is the side that matters most for the long-run frontier.

Thomas Kuhn’s old distinction between normal science and paradigm shifts is useful here. Normal science is the puzzle-solving work within a paradigm. It is recombinatorial, cumulative, and mostly successful. AI is exceptionally well-suited for normal science. It can search the literature, propose hypotheses within a known frame, and pattern-match across cases better than any human. The recent successes that look like AI accelerating science, including the protein-folding work of AlphaFold and the various materials-discovery cases, are largely cases where humans have already framed the problem cleanly and AI does powerful normal-science work within the frame. The wins are real. They are not the whole story.

What AI does not yet do, and what the developmental story suggests it may do less well as a co-product, is notice that the literature is asking the wrong question. Paradigm shifts are produced by people who feel a kind of discomfort with the existing frame that the frame itself does not generate. That discomfort is, in my reading of history, almost always a product of deep, embodied expertise. Novices do not produce paradigm shifts. The cohort that grew up with AI doing the framing for them will produce fewer of them than they otherwise would have.

The productivity boom can coexist with a slowdown in the rate of paradigm shifts. The two are measured on different timescales and by different instruments. We may notice the boom in five years and the slowdown in another 30.

The productivity boom from generative AI is real, and the gains for the people currently in a position to capture them are substantial. The trap is that the conditions producing the boom are not self-sustaining. The expert class that extracts value from these tools is a stock, not a flow. The flow that would replenish it is being slowed, in ways the current data are not designed to detect.

This is the diagnosis. The harder question is what to do about it: how schools, universities, professional training, research funding, and the design of the tools themselves might be arranged to keep building expertise rather than quietly spending it. I take that question up in the forthcoming companion piece.