Repaying the inheritance: How education and research policy can address AI’s borrowed expertise

In a companion piece, I made an argument that I find more worrying the longer I sit with it. The productivity boom from generative artificial intelligence (AI) is real, but it is being powered by people whose expertise was built before the tools existed. They are good at directing these systems because they spent years learning to tell a good answer from a confident-sounding wrong one, and that judgment was earned the slow way, by working through hard problems unaided and being corrected. The trouble is that the gains we can measure come mostly from routine work, the kind that is really information retrieval, while the work that defines genuine expertise lies elsewhere, in the framing of new problems and the creation of new knowledge, where AI helps least and the user’s own depth matters most.

What makes the situation precarious is that the conditions that built this expertise are being quietly removed. Juniors do less of the developmental work, firms hire fewer of them, and a research culture that once rewarded slow and idiosyncratic judgment is being nudged toward recombination. We are spending an inheritance of expertise that took decades to accumulate, and we are not replenishing it at the same rate. The cost does not show up in any of the numbers we watch. It is invisible until it isn’t, and by then it is expensive to repair.

That article was the diagnosis. We now turn to what to do, yet any solutions are unlikely to fully dissolve the underlying tension. There is a real conflict between what makes us productive this quarter and what builds capable people over a decade, and none of the proposals here make that conflict disappear. Even so, it is important to make the tension visible and to give schools, universities, professional bodies, and funders a vocabulary and a set of instruments for managing it, before the most consequential decisions get made for us by default.

There is a useful precedent here. When pocket calculators became cheap in the 1970s and 1980s, mathematics teachers faced a question that looks like the one we face now. The calculator could do arithmetic better and faster than any student. Why force students to do arithmetic by hand?

The answer that emerged, and that has held up reasonably well, was sequencing. Students learn arithmetic by hand first. They are tested without calculators in early grades. Calculators are introduced gradually, and only after the underlying operations have been internalized to the point where the student can recognize a wrong answer when they see one. The calculator becomes an instrument of speed and scale, not a substitute for the cognitive work that gives the student a sense for whether the output is plausible.

The AI case is structurally similar, but in three respects it is harder, and the policy response needs to be correspondingly stricter rather than looser.

First, calculators give correct answers within their domain. AI does not. A calculator that says seven multiplied by eight is 56 is right; a model that says a particular contract clause is unenforceable, or that a particular drug interaction is safe, may be confidently wrong, and only the user’s domain knowledge will catch the error. The user without that knowledge is worse off than a calculator user. They have a tool that produces plausible-sounding mistakes and no internal capacity to detect them.

Second, AI does not just execute the operation the way a calculator does. It chooses the framing. A student with a word problem and a calculator still must recognize that the problem requires subtraction rather than addition; the calculator does not do that for them. A student with a word problem and a language model is offered an answer that includes the framing as part of the package. Whether the framing is right, the user often cannot tell. This part of the work, the part that decides what operation is even being performed, is precisely the part that domain expertise builds and precisely the part that AI now does silently and on the user’s behalf.

Third, the convenience differential is much larger. A calculator saves a student seconds of arithmetic. A language model saves a student an entire essay, an entire memo, an entire research summary. The temptation to skip the developmental work is much larger because the work being skipped is much larger. The market and social pressures to reach for the tool are correspondingly stronger.

These three differences argue for an approach to AI in education that is, if anything, more conservative than the approach we eventually settled on for calculators—not prohibition, but careful sequencing, and with the bar for permitting unrestricted use set higher than it was for arithmetic. With that framing in mind, the policy agenda has four components: K-12 education, higher education and professional training, research and innovation funding, and AI design and procurement standards.

The K-12 question has three components that the public debate routinely conflates: whether students should use AI in their schoolwork, what students should be taught about AI, and how teachers should be prepared to handle both.

First, the principle is sequencing. In the early grades, AI tools should be largely absent from core academic work, because the cognitive operations trained at that stage—reading, writing, basic mathematics, basic reasoning—are the substrate on which all later learning depends, and the student should be required to develop that substrate unaided. In high school, structured and supervised use becomes appropriate after the underlying competence has been demonstrated without the tool, in the same way calculators are introduced for higher mathematics after arithmetic has been mastered.

Second, AI literacy belongs in the K-12 curriculum the way basic civics and basic financial literacy do. A defensible curriculum would cover four areas: how the systems actually work, how they fail, their social and economic effects, and the ethical issues their use raises. The point is to produce citizens who can reason about AI as a phenomenon, not just users who can prompt it.

Third, teacher training is the binding constraint. Neither the sequencing nor the literacy curriculum can be delivered by teachers who do not themselves understand the technology, and the federal and state investment required to close that gap is substantial and largely not yet committed. This is, in my view, the single highest-leverage education-policy investment available right now, and the one that current education-policy conversations are most reliably failing to take seriously.

Higher education raises a harder version of the same problem. University students are adults, the institutions are decentralized, and prohibition is impractical. The policy response must operate through credentialing rather than restriction.

The principle should be that for any credential whose value depends on the holder’s underlying capacity, the credentialing institution must be able to certify that the holder has, in fact, developed the underlying capacity unaided. This means AI-free assessment at the credentialing stages of university and professional training. A medical student who used AI to write all their clinical reasoning during training, and who is then licensed to practice based on those reports, is a public-health problem regardless of how productive AI made their education. The same applies, with different specifics, to law, engineering, accounting, and the other licensed professions, and increasingly to research credentialing as well.

Credentialing institutions, including universities, professional licensing boards, and accreditation bodies, should be making explicit and conservative decisions about what fraction of student work must be produced unaided for the credential to mean what it has historically meant. This decision is being made implicitly, accidentally, and inconsistently right now, by default, in ways that vary across institutions and that are not visible to the public that relies on the credentials. Making this decision more transparent is policy that does not require new legislation, only the willingness of accreditors and licensing boards to take a position and defend it.

A separate point applies to professional training inside firms. Apprenticeship phases in fields like consulting, law, finance, and medicine have historically been the period when junior professionals develop the judgment that distinguishes them from competent novices. The current default, which lets junior staff use AI throughout the apprenticeship phase and measures their output rather than their development, is hollowing this period systematically. The seniority-biased hiring patterns documented by multiple studies suggest that many firms are going further still, significantly reducing hiring at the entry level. Firms that care about their long-term pipeline should be carving out structured, AI-free developmental phases in early-career training, even at some cost to short-term productivity. Government and professional bodies can encourage this through the credentialing mechanisms above, and through procurement standards for the public sector, which is itself a major consumer of professional services.

Funding instruments at the National Science Foundation, National Institutes of Health, and their international counterparts should be examined for whether they reward genuine novelty or AI-assisted recombination. The same machinery that has produced extraordinary normal-science output is, by its design, mostly insensitive to the difference. Pilot review categories that explicitly weigh whether a contribution introduces a new frame, versus extending an existing one, would be a useful experiment. So would small-scale funding programs designed for the kind of risky, slow, idiosyncratic work that paradigm shifts have historically required, and that the current funding environment systematically underproduces. A study documenting how AI can enhance individual creativity but lead to homogenization across content offers, among other things, an early warning about what an AI-saturated research literature will look like if these instruments are not adjusted.

Pedagogical AI, in the sense of systems that scaffold learning rather than substitute for it, is a possible product. The market will not produce it on its own at the scale required, because the unit economics of frictionless answer delivery are better than the unit economics of productive struggle. Government procurement of AI systems for schools, universities, and training institutions can specify that the systems scaffold rather than answer, that they enforce productive struggle, and that they are evaluated on the development of the user rather than the user’s short-term satisfaction. This is an engineering problem. It becomes a tractable engineering problem at the moment a sufficiently large buyer specifies it as a requirement. Government is the buyer that can make that specification, and the moment to make it is before the educational technology procurement decisions of the next decade are baked in around the wrong default.

None of these directions are a complete answer, and none of them remove the underlying tension between short-term productivity and long-term cognitive development. The point is to make the tension visible and give institutions a vocabulary and a set of instruments for managing it. At present, the tension is mostly invisible, the instruments do not exist, and the developmental costs are being silently absorbed by a generation that will live with the consequences.

The policy window for protecting the pipeline is now. In 20 years, when the pipeline has already failed and the productivity gains have started to recede for the simple reason that there is no one left to direct the tools, the cost of repair will be much higher, and the institutional knowledge required to repair it will itself have thinned. The argument for acting now is not that the threat is certain. It is that the cost of acting now is low, and the cost of acting late is very high.

We have built a remarkable technology. The question is whether we can use it without quietly disinheriting the generation that should have been its next set of masters.