Hollowing the Top

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI documents the hollowing of the top from the inside: a technology executive who discovered that AI could perform many of the generative tasks he had spent years mastering, while the evaluative and directorial tasks became more central to his work. This is precisely the restructuring the hollowing-the-top concept predicts—not the elimination of the occupation but its reorganization around the components most resistant to automation. The senior practitioner’s task bundle shifts toward judgment, direction, and evaluation; the generative components that previously constituted the bulk of the work migrate to the machine.

The distributional asymmetry that Autor’s framework identifies is where the concept becomes uncomfortable for the cycle’s optimistic register. The extraordinarily well-positioned practitioner who experiences the restructuring as an elevation cannot be taken as representative. The junior lawyer whose brief-writing tasks are automated may not be promoted to strategic advisor; she may find that the generative rungs through which she was expected to develop the judgment that would eventually qualify her for the evaluative role have been eliminated—the developmental paradox operating at occupational scale.

The concept also reframes the recognition-theoretic dimension of the disruption. The social contract governing professional life rested on an implicit bargain: invest in education, develop specialized expertise, and the labor market will reward you with stable, well-compensated employment and social standing. For the generation that grew up under this contract, the prospect of hollowing the top represents not merely an economic threat but an existential one—the devaluation of an investment made in good faith under a promise that the social order has altered.

Origin

The concept emerges from Autor’s sustained engagement with what generative AI has done to the analytical boundary his 2003 framework established. The framework’s central prediction was confirmed across two decades of data: computerization automated routine tasks in the middle of the wage distribution while complementing non-routine abstract tasks at the top. The U-shaped employment growth this produced—job polarization—became one of the most robust empirical regularities in labor economics.

The arrival of large language models challenged this framework not by refuting its historical explanatory power but by demonstrating that the boundary it identified was not fixed. Autor’s response, articulated in his 2024 NBER paper and various public engagements, was to rework the framework rather than defend it unchanged. The new analysis distinguishes two kinds of non-routine abstract capability: the generative (first drafts of documents, initial code implementations, preliminary analyses) and the evaluative (judging whether the draft serves its purpose, assessing architectural appropriateness, integrating the output into a larger strategic context). AI can now perform the first category at professional quality. The second category remains with the human practitioner—but in many occupations, the generative tasks are precisely the tasks through which junior practitioners developed the evaluative capacity they would need at senior levels.

Hollowing the top is therefore not primarily a threat to senior practitioners, whose evaluative skills and contextual knowledge are complemented rather than substituted by AI tools. It is primarily a threat to the pipeline through which the next generation of senior practitioners develops. The professional pyramid—broad at the base of junior generative workers, narrow at the apex of senior evaluative partners—may flatten not because senior roles are eliminated but because the base is automated away before the practitioners occupying it have climbed the learning curve.

Key Ideas

The boundary shift. What counts as non-routine is itself a function of available technology; a task is non-routine not because of any intrinsic property but because no machine currently exists that can perform it. The boundary between routine and non-routine is therefore a moving frontier, and the story of technological progress is the story of that frontier’s advance. Generative AI has not merely pushed the frontier further along its previous trajectory; it has jumped it into territory the original framework had assigned to the permanent domain of human advantage.

Generative versus evaluative tasks. The most consequential internal distinction within the non-routine abstract category is between generative tasks (producing first instances of professional outputs) and evaluative tasks (judging, selecting, integrating, and directing those outputs). AI is substituting for the generative while complementing the evaluative. The senior lawyer whose brief-writing is delegated to AI becomes more valuable as a strategic advisor; the junior lawyer who expected to develop evaluative judgment through brief-writing may find the developmental pathway closed. Complementarity and substitution operate simultaneously within the same occupation, at different career stages.

The pipeline problem. Professional expertise has historically been transmitted through graduated exposure to generative tasks of increasing complexity under diminishing supervision. The resident who reads thousands of imaging studies develops the perceptual system that eventually supports complex diagnostic judgment. The junior associate who drafts hundreds of contracts learns the law through practice. When AI performs these developmental tasks before practitioners have had the opportunity to learn from them, the profession retains its non-routine components at the top but loses the developmental infrastructure through which new practitioners acquire the capacity to perform them. This is the developmental paradox operating at the occupational level.

Accountability as a constraint on substitution. In many high-stakes professional domains, AI substitution is constrained not by capability but by accountability. A large language model can perform a radiology read with comparable accuracy to a physician; but the physician who explains a diagnosis to a frightened patient, weighs risks and benefits against the patient’s specific values, and takes legal and ethical responsibility for a clinical decision is doing something that has no current AI equivalent. The relevant distinction may be shifting from cognitive (can the machine follow the rules?) to institutional (can the machine bear the responsibility?). Where accountability constraints are strong—medicine, law, engineering—hollowing the top may proceed slowly. Where they are weak, faster.

Political economy of credentialed displacement. The hollowing of the middle generated populist responses from displaced middle-class workers. Hollowing the top could generate a different political response: the frustration of a credentialed professional class whose substantial educational investment—often financed by debt—has been devalued by technological change. The political economy of a society in which large numbers of highly educated people feel their human capital has been arbitraged will not resemble the political economy of the previous era’s automation anxiety.