WORKING NOTE · v1 · 2026-06-11

Methods

Summary. Metro exposure = Σocc (exposure score × occupation's share of local employment), computed for six published occupation-level indices over BLS OEWS May 2025 employment data, reported as employment-weighted percentiles with full cross-methodology ranges. The timeline scenario is a separate, parameterized layer — never presented as data.

1. Data

Occupational employment per metro: U.S. Bureau of Labor Statistics, Occupational Employment and Wage Statistics (OEWS), May 2025 release — 393 metropolitan statistical areas (OMB Bulletin 23-01 delineations), detailed 2018 SOC. OEWS counts wage-and-salary jobs at the establishment's location; it excludes the self-employed, military, and most agriculture [1]. Population context: Census Population Estimates Program, Vintage 2025. Every source file's URL, vintage, license, and checksum is in the public provenance manifest.

2. The six indices

No single exposure measure deserves your trust; the spread across them does. Toggleable indices:

IndexWhat it measuresKey caveat
Eloundou et al. 2024 (OpenAI) [2]Share of an occupation's tasks where an LLM cuts completion time ≥50% (β = E1 + 0.5·E2; we average the human and GPT-4 ratings)GPT-4-era snapshot (early 2023); silent on augmentation vs substitution
Felten, Raj & Seamans 2023 [3]Relatedness of occupational abilities to language-modeling progressAbility-level ratings (mTurk); correlates with education generally
Microsoft applicability 2025 [4]Observed Bing Copilot usage mapped to work activities (200K conversations)Consumer assistant usage skews to writing/information tasks
Anthropic Economic Index 2026 [5]Observed Claude usage mapped to O*NET tasks/occupationsClaude users ≠ the workforce; usage ≠ displacement
Eisfeldt et al. 2024 [6]Generative-AI task exposure built for firm-value analysisDesigned for equity pricing, not employment
ILO Gmyrek et al. 2025 [7]GPT-4o task scores calibrated on 52,558 human judgments (ISCO-08)Reserved for non-US tiers (native ISCO); not in US consensus

Table 1. Prediction-style indices (rows 1–2, 6) rate what AI could do; usage-based indices (rows 3–4) measure what people actually do with AI today. Where they disagree, the band on every chart widens — that is the honest state of knowledge.

3. Normalization and aggregation

Raw index values are incommensurable (probabilities, z-scores, usage shares). Each index is converted to an employment-weighted percentile over national employment: a score of 70 means "more exposed than 70% of US workers." Metro exposure per index is the employment-weighted mean over the metro's detailed occupations (the method of Brookings' metro analyses [8]). The consensus is the unweighted mean of available index percentiles; the band is their min–max. O*NET-SOC detail codes are averaged to 6-digit SOC; suppressed OEWS cells are excluded and coverage is reported per metro as a grade (A ≥85% of area employment matched, B ≥70%, C below).

4. The scenario model

The timeline is governed by one equation — displaced(occ,y) = β·a·C·σ(k(y−T−L)) — where β is the occupation's exposed task share [2], a the automation share of AI use, C the adoption ceiling, σ the logistic curve with steepness k centered at arrival year T plus friction lag L. Layoffs are the displacement flow in excess of natural attrition r: if diffusion is slow enough, positions disappear through retirements rather than firings. Every parameter, its default, and its justification is exposed at /model; the defaults constitute the "median" preset. This layer is a scenario, not a measurement, and is visually segregated from data throughout the site.

5. Observed reality (the discipline)

Any scenario must pass through what has actually happened. Through early 2026: occupational churn at the aggregate level shows no relationship with exposure measures (Yale Budget Lab tracker [9]); yet employment for workers aged 22–25 in the most automation-exposed occupations has fallen 13–16% relative to peers (Brynjolfsson, Chandar & Chen [10]); and AI was the #1 stated reason in US layoff announcements for March–May 2026 (Challenger [11]). These facts are jointly compatible — an entry-level effect inside an aggregate null — and the default parameters reproduce them: modeled displacement through 2026 stays under 1.5% everywhere.

6. Validation

The build fails unless: (i) our Eloundou-β reproduction matches Brookings' published San Jose anchor (43% of workers in occupations with ≥50% tasks exposed) within ±2 points — current build: 42.9%; (ii) rank-order geography reproduces (San Jose most exposed large metro, Las Vegas bottom decile); (iii) every metro reconciles ≥50% of area employment; (iv) face-validity checks on the most/least exposed occupations pass; (v) every statistic traces to a manifest entry. Full report: validation report.

7. Limitations

Exposure ≠ job loss. Most indices are explicitly agnostic between augmentation and substitution. Threshold artifacts: the "share of workers with ≥50% of tasks exposed" stat is knife-edge sensitive — in Las Vegas, ~27K managerial jobs sit at β 0.48–0.49, which is why our threshold stat runs below Brookings' despite identical method; the primary metric is continuous and immune. OEWS exclusions: self-employed workers (disproportionately writers, designers, drivers) are invisible here. Crosswalk noise: managerial occupations map worst across classification systems. Vintage: one OEWS vintage per build; OEWS is not a time series and year-over-year comparisons are invalid.

References

  1. U.S. BLS, OEWS May 2025. bls.gov/oes
  2. Eloundou, Manning, Mishkin & Rock (2024). GPTs are GPTs. Science 384:1306–1308. doi:10.1126/science.adj0998
  3. Felten, Raj & Seamans (2023). How will language modelers affect occupations? arXiv:2303.01157; and (2021) SMJ 42(12). doi:10.1002/smj.3286
  4. Tomlinson, Jaffe, Wang, Counts & Suri (2025). Working with AI. arXiv:2507.07935. Data CC BY 4.0.
  5. Anthropic Economic Index (2025–26). anthropic.com/economic-index. Data CC-BY.
  6. Eisfeldt, Schubert, Taska & Zhang (2024). The labor impact of generative AI on firm values. SSRN 4436627.
  7. Gmyrek et al. (2025). Generative AI and jobs: a refined global index. ILO Working Paper 140.
  8. Muro, Methkupally & Kinder (2025). The geography of generative AI's workforce impacts. Brookings.
  9. Yale Budget Lab (2025–26). Evaluating the impact of AI on the labor market.
  10. Brynjolfsson, Chandar & Chen (2025). Canaries in the coal mine. Stanford Digital Economy Lab.
  11. Challenger, Gray & Christmas (2026). Monthly job-cut reports.