Demographics of AI-Exposed Workers

One-line

Workers in occupations most exposed to AI (per observed-exposure-measure) are older, more female, more white/Asian, much more educated, higher-paid, and less unionised than workers in unexposed occupations. They are not the demographic the prior bear-case projection (Acemoglu) predicted.

The pattern (per massenkoff-mccrory-labor-market-impacts-2026 Figure 5)

Top quartile of exposure vs zero-exposure workers, Aug–Oct 2022 CPS (just before ChatGPT release):

Dimension	Zero exposure	Top quartile	Difference
AI coverage %	0.0%	38.8%	+38.8 pp
Demographics
Age	41.0	42.9	+1.9
Female %	38.8%	54.4%	+15.5 pp
Hispanic %	24.8%	13.8%	−11.0 pp
White, non-Hispanic %	54.5%	65.1%	+10.6 pp
Black, non-Hispanic %	13.2%	9.7%	−3.5 pp
Asian, non-Hispanic %	4.7%	9.1%	+4.4 pp (≈2×)
Married %	44.6%	54.9%	+10.4 pp
Education
Less than HS	13.2%	2.3%	−10.9 pp
HS diploma	38.9%	17.7%	−21.2 pp
Some college / assoc.	30.0%	25.5%	−4.6 pp
Bachelor’s	13.3%	37.1%	+23.8 pp
Graduate degree	4.5%	17.4%	+12.8 pp (≈4×)
Labor market
Hours/week	37.5	38.7	+1.2
Hourly wage ($)	$22.23	$32.69	+$10.45 (+47%)
Union member %	11.7%	5.3%	−6.4 pp

See high-vs-low-exposure-worker-characteristics for the data citation.

The headline asymmetries

• Education is the strongest separator. Workers with graduate degrees are ~4× over-represented in the most-exposed quartile (17.4% vs 4.5%). Workers without a high-school diploma are ~6× under-represented (2.3% vs 13.2%).
• A 47% wage premium distinguishes exposed from unexposed work. The exposed group earns $32.69vs$22.23 per hour.
• The female over-representation (+15.5 pp) is large but lives alongside the education over-representation; the typical exposed worker is a college-educated woman in administrative, healthcare-records, marketing, or finance support work.
• Union membership halves in the exposed group (5.3% vs 11.7%), consistent with white-collar non-unionised work.

Why this matters

A direct contradiction (partial) with acemoglu-simple-macroeconomics

Acemoglu (2024) predicted the wage incidence of AI displacement would fall on low-education, native-born women in administrative and routine cognitive work. The Anthropic data shows the most-exposed workers are highly educated (37% Bachelor’s; 17% graduate degree) and higher-paid, not low-education. The female over-representation aligns with the prediction; the education and wage profile contradicts it.

The two are not strictly incompatible — Acemoglu’s measure was task-based from Eloundou’s earlier work and lacked usage weighting; Anthropic’s measure adds realized Claude deployment. The reconciliation may be that task-level theoretical exposure (Acemoglu) concentrates on routine clerical tasks performed by lower-education workers, but realized LLM usage (Anthropic) is happening disproportionately in higher-skill knowledge work because that’s where current LLM products are sold, used, and trusted. As the theoretical-vs-observed-capability-gap closes, the incidence could shift downward — or the gap could remain structurally uneven.

This is a live research disagreement, not a closed question. Either prediction could be retrospectively correct; current evidence sides with Anthropic’s empirical measure over Acemoglu’s theoretical projection.

A direct corroboration of the H2_u-curve-of-value labor-side story (sort of)

The labor-exposure profile produces a U-shape on a different axis from the value-stack U:

• Top of the exposure U: high-wage knowledge workers (programmers, analysts, lawyers, accountants).
• Bottom of the exposure U: manual-physical labor (cooks, mechanics, lifeguards) — protected by embodiment.
• Middle: semi-skilled service work, partially exposed.

If labor displacement eventually materialises (currently invisible per unemployment-did-exposed-workers), the burden lands on the upper half of the wage distribution — a notable inversion of prior automation waves (manufacturing automation hit the middle; offshoring hit the middle; AI hits the top).

Surplus distribution implication for capital-labor-divergence

If exposed labor is the already-higher-paid half of the workforce, AI-driven productivity gains in those occupations route to capital owners of the L4–L6 application layer (Cursor, Anthropic, etc.) rather than to the displaced workers themselves. The lower half is untouched in either direction — neither displaced nor lifted.

The “AI hurts everyone equally” framing is wrong on the labor side. So is “AI displaces the most vulnerable first.” Current evidence: AI exposure correlates positively with skill, education, and pay.

Caveats

• Snapshot pre-treatment. The table is Aug–Oct 2022 — before ChatGPT release. Post-2022 demographic shifts within these groups are not captured here.
• CPS occupation-level only. No within-occupation skill-level decomposition. Within “Computer programmers” there are juniors and seniors; the table cannot show whether one is more exposed than the other.
• US data. Cross-bloc comparison is absent; the demographics of exposed workers in EU or East Asian labor markets is unknown.

Related

• massenkoff-mccrory-labor-market-impacts-2026 — source.
• high-vs-low-exposure-worker-characteristics — data citation.
• observed-exposure-measure — the measure these demographics are scored against.
• most-exposed-occupations — the occupational composition of the top-quartile group.
• acemoglu-simple-macroeconomics — the prior prediction this evidence partially contradicts.
• capital-labor-divergence — the surplus-distribution lens; this evidence updates the labor-incidence prediction.
• H2_u-curve-of-value — labor-side U-shape on a different axis from the value-stack U.
• ai-young-worker-hiring-slowdown — the entry-level signal that may indicate where the displacement actually lands.
• task-based-framework — Acemoglu’s macro frame; this evidence sits inside it.
• diffusion-vs-innovation — why the realized-incidence differs from theoretical projection.
• fusion-skills — the labor-side capability question for the cohort most exposed.
• displacement — the concept this demographic profile is the empirical anchor for.