Published in the Journal of Economic Perspectives 29(3), 3–30, by David H. Autor (MIT). It is the framework paper behind Autor's later Applying AI to Rebuild Middle-Class Jobs, and the standard citation for why automation has not historically produced net job loss.
The central claim
Autor's diagnosis of the discourse: "journalists and even expert commentators tend to overstate the extent of machine substitution for human labor and ignore the strong complementarities between automation and labor that increase productivity, raise earnings, and augment demand for labor."
The paper's positive account is that automation "complements labor, raises output in ways that lead to higher demand for labor, and interacts with adjustments in labor supply." Technology changes "the types of jobs available and what those jobs pay" without necessarily reducing their number.
Polanyi's paradox
The bound on substitution is epistemic rather than technical. Computerization substitutes for tasks that can be specified in explicit rules, which "has led to a substantial decline in employment in clerical, administrative support, and to a lesser degree, in production and operative employment."
"But the scope for this kind of substitution is bounded because there are many tasks that people understand tacitly and accomplish effortlessly but for which neither computer programmers nor anyone else can enunciate the explicit 'rules' or procedures."
Autor names this Polanyi's paradox, after Michael Polanyi's 1966 observation: "We know more than we can tell." His examples: "when we break an egg over the edge of a mixing bowl, identify a distinct species of birds based on a fleeting glimpse, write a persuasive paragraph, or develop a hypothesis to explain a poorly understood phenomenon, we are engaging in tasks that we only tacitly understand how to perform."
The resulting prediction: "the tasks that have proved most vexing to automate are those demanding flexibility, judgment, and common sense — skills that we understand only tacitly."
This is the claim most directly tested by large language models, which acquire capability from examples rather than from enunciated rules — the paradox constrains rule-based programming specifically, and machine learning was designed to route around it.
The O-ring argument
The complementarity mechanism is formalized through Kremer's (1993) O-ring production function, in which "failure of any one step in the chain of production leads the entire production process to fail. Conversely, improvements in the reliability of any given link increase the value of improvements in all of the others."
Autor's intuition for it: "if n−1 links in the chain are reasonably likely to fail, the fact that link n is somewhat unreliable is of little consequence. If the other n−1 links are made reliable, then the value of making link n more reliable as well rises."
Applied: "when automation or computerization makes some steps in a work process more reliable, cheaper, or faster, this increases the value of the remaining human" tasks. Where inputs "each play essential roles… improvements in one do not obviate the need for the other," so "productivity improvements in one set of tasks almost necessarily increase the economic value of the remaining tasks."
This is the analytic core of the complementarity claim, and the mechanism displacement arguments must defeat: they generally do so by asserting that AI eventually covers the remaining links rather than raising their value.
Polarization
Autor documents the "polarization" of the labor market — "wage gains went disproportionately to those at the top and at the bottom of the income and skill distribution, not to those in the middle" — and offers evidence for it, while arguing "this polarization is unlikely to continue very far into the foreseeable future."
The paper's final section turns to what advances in artificial intelligence imply for the framework.
Relationships
- depends-on: AI Labor Disruption — the framework the displacement thesis argues against
- related: Applying AI to Rebuild Middle-Class Jobs — Autor's later application of this framework to AI specifically
- related: David Autor, AI and Productivity, Remote Labor Index: Measuring AI Automation of Remote Work (Mazeika et al., 2025)