Abstract

We propose a methodology for using online job posting data to measure how generative artificial intelligence is restructuring US knowledge work, and we articulate the patterns that distinct theoretical hypotheses predict empirical implementations of this methodology would reveal. The design constructs an occupational exposure measure from a task-level mapping between O\*NET task descriptions and current AI capability, classifies postings by exposure quintile, and applies a difference-in-differences specification with the November 2022 release of large language models as the focal event. Three outcome margins receive specific attention: the volume of postings within and across exposure categories; the within-occupation composition of skill requirements; and the within-occupation distribution of posted wages. For each margin, we develop a measurement procedure, specify the estimating equation, and articulate the empirical patterns that distinct hypotheses—substitution, complementarity, restructuring, reorganization, and null effect—would predict. We embed the design within an explicit task-based model that derives testable implications from microfoundations, discuss the construct validity of the exposure measure under shared-prior bias among raters, provide statistical power calculations anchored to published within-occupation residual variances, develop a ten-item pre-registration protocol that disciplines specification search and commits to multiple-testing correction and code-deposit, demonstrate the procedure end-to-end with a synthetic worked example under each hypothesis, and bound the residual identification limits arising from joint macroeconomic shocks contemporaneous with the focal event. We do not present empirical estimates from a particular dataset, and we are explicit that the design recovers a cross-sectional differential rather than a clean causal effect. The paper specifies the methodology that an empiricist with access to comprehensive job posting data could implement, and articulates the interpretations that empirical results would support. By separating design specification from implementation, the paper aims to support a more disciplined empirical literature on a question whose answer carries substantial implications for the measurement of human capital, the design of education policy, and the projection of long-run labor market outcomes.