The Missing First Rung: Aggregate Resilience and Entry-Level Compression in the AI Labor Market

Abstract

Public discussion of artificial intelligence and employment has split between a jobs apocalypse and an AI jobs boom. Both claims are easier to make than to count because creation is often measured with broad infrastructure employment while displacement is restricted to explicitly attributed layoffs. I construct a symmetric occupation-industry ledger from Current Population Survey microdata covering January 2015 through August 2026. The primary universe fully excludes electric-power cells; power appears only in a separate sensitivity analysis. I also formalize a demand-expansion mechanism in which AI can increase output while reducing the junior labor required per unit of output. Aggregate results show resilience rather than collapse, but not a model-invariant boom. In an observed-use design, the post-ChatGPT aggregate high-exposure employment gap ranges from minus 0.47 to plus 0.34 percentage points across four counterfactuals. Its first-year gap lies 0.009 points above the maximum of five post hoc fully pre-pandemic placebo windows, but the five-window comparison has a minimum attainable randomization p value of 0.167; the gap also lies inside the wider distribution formed by eight pre-event breaks, three of which overlap the pandemic. The latest broad occupation-industry balance is 0.52 to 0.97 million jobs above trend, but its post-event mean ranges from minus 0.25 to plus 0.12 million. For workers age 22 to 25, high observed-exposure employment is 2.19 to 2.58 percentage points below trend across specifications; the primary estimate is minus 2.58 points. Their probability of moving from nonemployment into any job is 4.30 points below trend in the primary model and remains negative under all four counterfactuals. I read this as first-rung compression within a broader and partly model-dependent youth employment slowdown, not as a pure AI substitution effect. Ages 26 to 35 have negative average stock gaps in three of four counterfactuals and negative cell-ledger balances in all four, so they are not a clean unaffected control. The bottom line is aggregate resilience with a pronounced first-rung compression whose exact causal AI share remains unidentified.

Author abstract from the research manuscript. Typography and line breaks have been normalized; the linked source may show a different manuscript version.