AI-Wiki

Labor market impacts of AI: A new measure and early evidence

5 min read

Labor market impacts of AI: A new measure and early evidence

The methodological primary source behind the “observed exposure” labor measure the wiki has been citing second-hand. Published by Anthropic (5 March 2026) as part of the Anthropic Economic Index program; authors Maxim Massenkoff (Naval Postgraduate School) and Peter McCrory (Anthropic). It is the study Brovich (AWS) described on stage (“Anthropic from March of this year… observed vs theoretical exposure”), and the analytic backbone for several claims on ai-employment-effects.

Key findings

  • A new measure — observed exposure. Combines (1) theoretical LLM capability (Eloundou et al. 2023 task β-ratings — can an LLM make the task ≥2× faster) with (2) real-world usage from the Anthropic Economic Index, weighting automated (rather than augmentative) and work-related uses more heavily, aggregated from O*NET’s ~800 occupations’ tasks up to the occupation level. The automation/augmentation weighting is the methodological heart — it is what separates displacement risk from mere exposure.
  • AI is far from its theoretical ceiling. Actual coverage is a fraction of what’s feasible. Tasks rated fully feasible (β=1) account for 68% of observed Claude usage; β=0 tasks just 3% — usage concentrates where capability is highest, but most theoretically-automatable work is not yet being done by AI at scale.
  • Most-exposed occupations. Computer Programmers (75% coverage), Customer Service Representatives, Data Entry Keyers (67%), financial analysts. 30% of workers have zero coverage (Cooks, Bartenders, Lifeguards, Dishwashers, Motorcycle Mechanics, Dressing Room Attendants) — tasks too infrequent in the data to meet threshold.
  • Exposure tracks (weakly) with lower projected growth. Regression at the occupation level (employment-weighted): every +10pp coverage → −0.6pp in the BLS 2024–2034 growth projection. Slight but real validation against independently-derived analyst estimates — and notably the Eloundou measure alone shows no such correlation, evidence the usage-weighting adds signal.
  • Who the exposed are. Top-quartile-exposure workers (measured Aug–Oct 2022, pre-ChatGPT) are 16pp more likely female, 11pp more likely white, ~2× as likely Asian, earn 47% more, and are far more educated (graduate degrees: 4.5% of unexposed vs 17.4% of most-exposed — ~4×). This inverts the older “AI hits low-skill first” narrative.
  • No systematic unemployment effect yet. Difference-in-differences (top-quartile vs zero-exposure, CPS) since ChatGPT: the gap change is small and statistically indistinguishable from zero. The framework could detect a ~1pp differential increase; a “white-collar Great Recession” (top-quartile unemployment doubling 3%→6%) would be visible.
  • Suggestive young-worker hiring slowdown. Using the CPS panel’s job-start rate for ages 22–25: entry into the most-exposed occupations falls ~½pp; the post-ChatGPT estimate is a ~14% drop in the job-finding rate vs 2022 (barely statistically significant; no such drop for >25). Echoes Brynjolfsson et al.’s 6–16% employment fall for 22–25-year-olds in exposed occupations (they attribute it to slowed hiring, not separations) — here corroborated on a different dataset, with caveats (non-employed young workers may stay in school, switch jobs, or exit the labor force rather than appear unemployed).

Methodological stance (worth importing into the wiki)

  • Humility about forecasting. Past approaches over-predicted: offshorability flagged ~¼ of US jobs as vulnerable, most of which kept healthy growth; government occupational forecasts “added little predictive value beyond linear extrapolation.” The paper’s aim is to establish a measure now, before effects emerge, and revisit periodically — so future analyses separate signal from noise rather than rationalising post-hoc.
  • AI is more like the internet / China-trade than COVID. Effects are likely gradual and confounded by the business cycle and trade policy — “most useful when the effects are ambiguous,” not when they’re unmistakable.
  • Unemployment chosen as the priority outcome (over job-postings or occupation-mix shifts à la Gimbel et al.) because it most directly captures harm: a decline in postings for one exposed role may be offset by openings in a related one, but an unemployed worker wants a job and hasn’t found one.

What was actually ingested

The full 17-page PDF (the “Nowcasting” working version, v16, of the published note). Figures (1–7) were read from their captions and the surrounding text, not reproduced. The companion dataset (observed coverage at task and job level) is published at huggingface.co/datasets/Anthropic/EconomicIndex.

Why this matters to the wiki

This is the primary methodology the wiki’s labor-effects claims have leaned on through secondary citation (the AWS keynotes, the observed exposure framing on ai-employment-effects). Ingesting it: (a) gives the no-unemployment-yet / young-hiring-slowdown finding a citable primary; (b) supplies the automated-vs-augmentative weighting as a measurement principle that sharpens automation-vs-augmentation (exposure ≠ displacement risk; the weighting is the difference); (c) anchors the “capability ≫ adoption” gap (β=1 tasks = 68% of usage, yet most automatable work isn’t automated) that complements GDPval’s capability-as-leading-indicator framing from the usage side.

Linked entities and concepts

Relationships

Relationships

Part of

  • Anthropic Economic Index — introduces the labor-impact (observed-exposure) measure built on the Economic Index usage data; the analytic complement to the AEI usage reports

Supports

Published by

  • Anthropic — Anthropic research note; part of the Anthropic Economic Index program; data on Hugging Face (Anthropic/EconomicIndex)

Graph View

Backlinks