The question of urgency has already been argued in this edition on epistemic grounds: convergent AI output produces a form of influence with no historical precedent in its combination of scale, speed, and invisibility. The question of method has been argued on grounds of institutional practice: what makes a claim, including this Institute’s, worth believing. This essay argues a narrower and more specific claim, one that does not depend on either of the other two: that the current moment is technologically distinct from prior waves of automation in a way that can be stated precisely, and that the distinction bears directly on the tradition this Institute draws from.
What makes an era-defining technology
Bresnahan and Trajtenberg (1995) proposed the concept of a general-purpose technology to explain why whole eras of economic growth appear to be driven by a small number of technologies — the steam engine, the electric motor, the semiconductor — rather than by innovation spread evenly across the economy. What distinguishes a general-purpose technology from an ordinary one, in their account, is pervasiveness across sectors, continued potential for improvement, and what they call innovational complementarities: the property that improvements in the technology make innovation in the many downstream activities that depend on it more productive, not merely more automated.
That artificial intelligence plausibly qualifies as a general-purpose technology in this sense is not, at this point, a controversial claim, and this essay does not attempt to establish it. The more specific and more useful question is what kind of general-purpose technology it is — which activities it substitutes for and which it complements — because the answer to that question has, for the last several decades, been remarkably stable, and it has just changed.
What the last wave actually did
Autor, Levy, and Murnane (2003) gave that stability its most precise empirical account. Studying the computerization of the American workplace between 1960 and 1998, they found that computer capital substituted for labor specifically in routine tasks — those that could be accomplished by following explicit, codifiable rules — while it complemented labor in non-routine cognitive and interpersonal tasks: problem-solving, communication, and judgment exercised under conditions too variable to reduce to a fixed procedure. Computerization did not merely fail to threaten these capacities. It increased the market’s demand for them, accounting, by the authors’ estimate, for a substantial share of the shift in relative demand toward workers who could exercise them.
This finding did more than describe a labor market. It supplied, for the length of a generation, an implicit answer to the question of where human comparative advantage would continue to reside as automation advanced: not in calculation or procedure, which machines would increasingly perform, but in reasoning, judgment, and communication — the capacities a machine following explicit rules could not substitute for because they could not be reduced to explicit rules in the first place. A great deal of educational and economic policy, over the same decades, was built on some version of this assumption, often without stating it as one.
What changed
Eloundou, Manning, Mishkin, and Rock (2024), in a large-scale study of task exposure to large language models published in Science, found a pattern that inverts the one Autor, Levy, and Murnane documented rather than extending it. Exposure to language-model capability, in their analysis, is concentrated not in routine, codifiable work but in higher-wage occupations characterized by exactly the non-routine cognitive and communicative tasks that the prior computerization wave had left to humans, and had made more valuable by leaving to them.
This is the precise sense in which the current moment differs from the last one, and the difference can be stated without appeal to speculation about future capability: the technology that previously served as a general-purpose complement to reasoning and communication has been followed by one whose most direct effects fall on reasoning and communication themselves. The distinction is not that machines have never touched cognitive work before — calculators, spreadsheets, and word processors did so for decades, as instruments a person used to execute an argument or a calculation they had already formed. The distinction is that a system capable of originating the argument, not merely executing one already specified, is a categorically different kind of instrument, and it is this capacity — to produce the position, not merely process it — that the studia humanitatis was built specifically to cultivate in the human being who would otherwise have to originate it alone.
Four theses
Thesis 1. Prior general-purpose technologies, including the computerization wave documented by Autor, Levy, and Murnane (2003), substituted for routine tasks and increased the relative economic value of non-routine cognitive and communicative work, establishing reasoning and communication as the durable human complement to automation for roughly four decades.
Thesis 2. Generative artificial intelligence is the first widely deployed general-purpose technology to directly target non-routine cognitive and communicative tasks, a pattern documented empirically in large-scale task-exposure research (Eloundou et al., 2024) that inverts, rather than extends, the labor-market pattern established by the prior computerization wave.
Thesis 3. This inversion is a difference in kind, not merely in degree, from previous waves of automation, because it removes the tacit assumption — embedded in several decades of labor-market adaptation and educational policy — that reasoning and communication constituted stable ground on which human comparative advantage would continue to rest as automation advanced.
Thesis 4. Because the studia humanitatis was built specifically to cultivate reasoning and rhetorical communication as capacities requiring deliberate formation, this technological inversion is not incidental to the humanist tradition's present relevance; it is the precise condition under which the tradition's founding claim — that these capacities must be cultivated, not assumed — becomes newly testable rather than merely of historical interest.
This essay’s claims are subject to challenge under the terms of the Institute’s Disputation Protocol. Editorial disclosure, including this publication’s AI review process, appears in the closing section of this edition.
References
Autor, D. H., Levy, F., & Murnane, R. J. (2003). The skill content of recent technological change: An empirical exploration. The Quarterly Journal of Economics, 118(4), 1279–1333.
Bresnahan, T. F., & Trajtenberg, M. (1995). General purpose technologies ‘Engines of growth’? Journal of Econometrics, 65(1), 83–108.
Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306–1308.
Disputation
This argument is open to challenge.
HumanitAI treats publication as the beginning of scrutiny rather than its end. Factual correction, counter-evidence, competing interpretation, and argument are welcome.
Claims that do not survive challenge should be revised or withdrawn with a visible record of what changed and why.
The goal is not to win an argument. It is to discover which claims survive it.