How we project the labor gap.

Two anxieties dominate talk about the American workforce, and they point in opposite directions: that the country is running out of workers as the baby boom retires and immigration tightens, and that it will soon have too many as AI automates their jobs. We argue the honest answer is both — and that the two are features of the same labor market seen along different margins.

The trouble is that a "labor shortage" has long been suspect among economists. As Freeman (2006) put it, in a market a shortage is not a primitive fact but a statement about prices: if employers cannot find workers at the wage they wish to pay, the textbook remedy is to pay more. A shortage claim is therefore empty unless it says how far wages would have to rise to clear the market. Standard forecasts (BLS Employment Projections, the state agencies' Projections Central) project employment — an equilibrium quantity, the joint product of supply and demand — and so cannot say whether a projected gap would be erased by a one-percent raise or survive a fifty-percent one. We build the wage mechanism into the definition precisely to answer that objection.

American Community Survey microdata (2010–2024, via IPUMS) drives the demographic supply model. Pooled across fifteen years, the ACS lets us estimate, by metro and occupation, four transition rates: occupational retention as workers age, the rate at which young cohorts enter each occupation, net internal migration, and immigrant inflows. The 2024 ACS also anchors the equilibrium with each cell's 2024 employment and mean wage. We identify 271 metropolitan markets — 264 from the IPUMS met2013 variable plus seven recovered through a population-weighted PUMA-to-metro apportionment (Billings, Missoula, Great Falls, Manchester–Nashua, Sioux Falls, Rapid City, Burlington) — covering roughly 70% of employment, with each state's non-metropolitan remainder carrying the rest.

The Autor–Thompson (NBER w33941) AI exposure measure is the novel input. In their expertise framework, what AI does to an occupation depends on which tasks it automates. Automating an occupation's inexpert tasks concentrates the remaining work in fewer, more expert hands — wages rise, employment falls (AI substitutes). Automating its expert tasks commoditizes the work — wages fall, employment expands (AI complements). The sign of the employment effect is thus occupation-specific, not a universal property of "automation." We map their estimates onto our 94 detailed occupations (25 expertise-raising, 17 lowering, 52 essentially unaffected) and consider three deployment intensities: baseline (shifts not yet realized at scale), gradual, and transformative (calibrated to the upper end of their exposure estimates — a deliberately aggressive bound, not a forecast).

State workforce-agency demand projections (Projections Central), benchmarked to the national BLS Employment Projections for 2024–2034 and allocated to metros with leave-one-out Bartik shift-share instruments. A Current Population Survey occupation-to-occupation transition matrix (2010–2024) lets displaced workers re-queue into other occupations they are educationally qualified for.

Supply is pure demography. A worker's occupational employment a decade hence is the sum of three flows — retention, entry, and migration — applied to the 2024 stock, using no information about wages or where labor is demanded. That independence is what gives the eventual reconciliation its content. Under the baseline, immigration contributes about 5.6 million workers over the decade.

Demand is a constant-elasticity inverse demand curve in each market–occupation cell, with own-wage elasticities calibrated to the labor-demand literature (between 1 for the most elastic occupations such as food service and farming, and 3 for the least, such as healthcare and legal). AI enters here, shifting the 2034 demand curve inward in expertise-raising occupations and outward in expertise-lowering ones.

Who works where is a discrete-choice (logit) model: within each market-and-education cell, workers sort across occupations on wages and a non-wage amenity, recovered by the Berry (1994) inversion so the model exactly reproduces each metro's observed 2024 occupational structure. Wage-sensitivity declines with education (less-educated workers, with fewer options, reallocate more readily). The outside option of non-employment lets the labor force itself expand and contract with wages.

Equilibrium is the wage vector at which supply and demand balance in every cell, found market by market by a tâtonnement in log wages. We solve at three geographic resolutions and report all three; our preferred specification, the Composite Spatial Equilibrium (CSE), solves each of the 271 metros as an independent market and adjoins a separate equilibrium for each state's non-metropolitan remainder, so the markets cover essentially the entire workforce without double-counting any worker.

At the 10% ceiling we project a baseline shortage of 13.1 million workers by 2034 — concentrated in management (2.0M), healthcare practitioners (1.6M), computer and mathematical (1.2M), education (0.9M), and transportation (0.9M); every one of the 22 major groups shows a positive shortfall. It is not a knife-edge artifact of the ceiling: the schedule falls smoothly from 16.3M at 5% to 5.3M at 30%.

How much does geography matter? Solving only the 271 metros (omitting the ~30% of employment outside them) yields 10.7M; solving whole states, which lets labor reallocate freely within a state, yields 12.5M; the CSE yields 13.1M. We read the three as a bracket around the headline, not as rivals, and prefer the CSE because it covers essentially everyone while conservatively assuming workers do not chase wages across metro lines over the decade.

AI does not erase the shortage — it rotates the problem. As deployment intensifies from baseline to transformative, the projected shortage at the 10% ceiling actually falls (13.1 → 10.5 → 9.9M), because AI shifts demand inward in the most exposed occupations; meanwhile re-queued involuntary unemployment climbs (1.4 → 3.1 → 8.4M). Netting out the mechanical floor, AI-attributable unemployment under the transformative scenario is 6.9 million on the preferred re-queued measure — up to 10.8 million if displaced workers cannot reallocate at all.

The mechanism's signature is that employment and wages can move in opposite directions within the same occupation. Under the transformative scenario, expertise-raising occupations shed workers as wages rise (Business & Financial −44% employment / +13% wage; Computer & Mathematical −33% / +10%; Legal −20% / +6%); expertise-lowering occupations expand as wages fall (Farming +21% / −6%; Production +15% / −5%; Transportation +12% / −4%). The demographic shortage and the technological displacement coexist because they fall on different occupations.

A distinctive feature of the framework is that immigrants appear not only as suppliers of labor but as consumers of it, through the local-demand channel documented by Hong & McLaren (2015) and Cortés (2008). This bidirectional treatment produces a result one-sided accounting cannot. Eliminating immigration over the decade widens the shortage (13.1 → 13.9M) — exactly as the "immigrants fill labor gaps" intuition predicts — but the same policy shrinks AI-driven unemployment (6.9 → 5.6M), because the workers not admitted do not generate the local demand whose later exposure to AI would have put jobs at risk. The two headline numbers move in opposite directions under the same policy. A model that treats immigrants as a pure supply shock sees only half the trade.

The entire equilibrium is denominated in constant 2024 dollars, with no aggregate price level. This is deliberate. The object of interest is a set of relative quantities — how many workers each occupation supplies and demands at a given configuration of relative wages — and the wage ceiling and floor are naturally real (the ceiling is a decade of real wage growth). Introducing an aggregate price level would scale all nominal wages together and wash out of the equilibrium conditions symmetrically: it would change the units we report wages in, but not the quantities that define shortages and unemployment. All wage bounds on this site are real, 2024-dollar magnitudes.

This matters for one chart. Where the landing page shows an inflation or price-side line alongside the wage increase needed to clear the market, that line is not a model output. It is an illustrative reduced-form passthrough — labor share × net wage pressure, spread over the ten-year horizon — that the reader can vary. The GE model says how far wages must move; it does not, by construction, say anything about the price level. We keep the two clearly separated.

We test the supply model out of sample by estimating its cohort rates on data through 2014 and projecting to 2024, where we can check it against what actually happened: it achieves a log RMSE of 0.32 across metro-by-occupation cells, against 0.47 for a random-walk benchmark — a third better. One correction materially shaped the headline: the new-entrant rate, estimated over a window ending in the pandemic year 2020, is biased downward by the 2020 collapse in youth employment; re-estimating over the COVID-free 2010–2019 window raises it from 0.76 to 0.90 (confirmed directly in the microdata) and lowers the projected shortage by about 3.4 million. We use the COVID-free window throughout.

The headline is nearly invariant to doubling the wage-elasticity of occupational choice and to removing the amenity shrinkage; it is, as any such model is, materially sensitive to the demand elasticity (interior to our literature-calibrated bounds). The one bias we cannot eliminate — that the agency demand projections may already embed suppressed shortages — works to make our headline a lower bound. The full pipeline runs end to end from public inputs (ACS, Projections Central, BLS, CPS, Census geography).