The displacement curve — what AI absorbs, and what it does not

What the literature actually says

The 2023 Goldman Sachs estimate that generative AI could expose ~300 million full-time-equivalent jobs to automation globally is the most-cited number, but the substance lies in the breakdown. Administrative support, legal, architecture and engineering, business and financial operations, life-and-physical-and-social-science, and management are the highest-exposure categories. Construction, installation, maintenance, building and grounds cleaning, food preparation, transportation, and personal-care services are the least-exposed.

The pattern is consistent across other 2023–2026 studies — ILO, OECD, McKinsey, the IMF, and the Brookings working papers all map cognitive and information-processing tasks as the displaced category, with embodied and care work as the residual.

The historical break

Every prior automation wave displaced manual or physical labor and was absorbed (over a generation, painfully) by the expansion of cognitive and service work. The mid-century moves out of agriculture absorbed into manufacturing. The late-century moves out of manufacturing absorbed into information work, knowledge professions, and services.

The current wave inverts that direction. The expansion category that absorbed previous displacements is itself the category being absorbed. There is no obvious next sector to absorb the displaced cognitive workforce at scale, because the work AI cannot do is, broadly, the work the displaced cohort has spent two generations being trained out of.

The implicit policy answer is dependency

The default response from international institutions — including the World Economic Forum's published "you will own nothing and be happy" framing — is to manage displacement through redistribution of AI-generated economic surplus to the displaced cohort as recipients. UBI variants, expanded social-welfare programs, and shareholder dividends from sovereign AI stakes are the mainstream proposals.

This is administration of dependency at population scale. It depends on the indefinite competence and good faith of the administering institutions. It also produces, at scale, the psychological profile of a population without work it is suited to do — which the wellbeing literature (the cubicle cost) describes in unflattering terms.

The reallocation alternative

The argument of this site is that the same productivity surplus, deployed differently, can fund the inverse move: cognitive work returns to the machines that do it without exhaustion; human bodies return to the work they were built for, funded by the surplus AI generates elsewhere in the economy. The displaced cohort is the expanding cohort, not the recipient cohort. Dependency is replaced by labor.

The agricultural absorption capacity is not theoretical — orchard, polyculture, and food-forest systems are documented to require roughly an order of magnitude more labor per acre than monoculture grain or row-crop systems. The labor is graded across abilities. The supporting infrastructure (processing, distribution, regional warehousing) extends the absorption beyond the field itself.

The question is not whether AI will displace cognitive labor. The literature converges on that. The question is what the displaced labor is asked to do next.