Does automation raise or lower the skills that remaining work demands?

Autor and Thompson argue that whether automation adds to or subtracts from the value of remaining labor "depends on whether removing tasks raises or reduces the expertise required for remaining non-automated tasks." The premise is that one task can be expert in one occupation and inexpert in another, so the same automation can replace experts in some occupations while augmenting expertise in others. To test this, they develop "a novel, content-agnostic method for measuring job task expertise" and apply it to occupational expertise demands over four decades. Their headline result is that automation "has raised wages and reduced employment in occupations where it eliminated inexpert tasks, but lowered wages and increased employment in occupations where it eliminated expert tasks."

The mechanism is what the authors call occupational task bundling. Removing inexpert tasks from a bundle raises the expertise the remaining work requires, which lifts wages but "reduces the set of qualified workers." Removing expert tasks lowers the requirement, which cuts wages but "permits the entry of less expert workers." The two channels run in opposite directions, and the employment effect is the one the authors stress: it is "distinct from—and in the case of employment, opposite to—the effects of changing task quantities." The framework is offered as a resolution of a familiar puzzle, why routine task automation has tended to lower employment while often raising wages in routine task-intensive occupations.

Against the nearest notes, this gives a direction to the scarcity argument in What makes accountable judgment scarce when AI cognition is cheap?. That note holds that accountable judgment becomes the scarce asset when cognition is cheap, and that institutions shape the outcome more than capability does. Autor and Thompson specify when scarcity value rises or falls: it depends on which tasks leave the bundle and what expertise they carried. Both treat occupations as bundles, but the paper's are defined by expertise requirements and the note's "governance bundles" by responsibility. Does concentrated AI exposure enable workers to adapt and reallocate? measures how exposure is spread across an occupation's tasks and finds that concentration allows reallocation to untouched work. The expertise framework asks a different question. Concentration tells you whether workers can move, while the expertise of what remains tells you what they earn and how many can qualify. The worker-level claim in Does AI turn freelance work into validation instead of creation? names a different channel. Autor and Thompson's "entry of less expert workers" concerns who can enter an occupation once requirements fall, not whether incumbents keep learning on the job. The two could compound, but the excerpt does not model skill formation.

The excerpt is the abstract alone. It names no data source, occupations or estimates, and it never mentions AI. The subject is task automation in general, measured through "job task removal and addition" over four decades. It therefore does not establish what generative AI removes from any given occupation, or whether the expertise of the work left behind rises or falls. The sign of the effects is therefore an empirical question the full paper would have to answer. At the strength the abstract allows, the implication is a question to put to any AI deployment: which tasks it takes out of an occupation's bundle, and whether the work left behind demands more or less expertise than before.