AI and robotics
are opposite stories.
Most reporting on automation and jobs collapses two distinct pressures into one panic. This dashboard separates them. AI hits knowledge work in coastal metros. Robotics hits physical work in Heartland manufacturing zones. The middle, in-person services and skilled trades, is the most insulated category in the US economy.
Three peer-reviewed exposure measures (Eloundou, Felten, Yale Budget Lab), BLS 2024–34 projections, and the seven head-to-head occupation pairs that show the structure. Search any job, plot the catalog, or compare two sides of the inversion.
Most public reporting on AI and jobs collapses three different questions into one number.
Exposure asks what fraction of an occupation's tasks an AI system could perform. A ceiling. Displacement asks how many of those tasks an employer will actually substitute capital for labor on, after weighing cost, regulation, and adoption frictions. Augmentation asks whether the technology makes remaining workers faster or enables new tasks.
Goldman's "300 million jobs exposed globally" is exposure. Acemoglu's "0.06 percent TFP growth per year over a decade" is a macro outcome. PwC's "56 percent wage premium for AI skills" is augmentation. None of them is wrong. All of them describe different layers. This dashboard keeps them separate.
The single most under-reported finding in the US labor data through April 2026 is an inversion. Every prior wave of automation hollowed out the middle of the wage and education distribution. Generative AI inverts that. Felten's AIOE, Eloundou's β scores, and the Yale Budget Lab harmonization all show exposure rising with credential, with bachelor's-and-above the most exposed. Robotics has not gone away. It is still the dominant pressure on warehouse, manufacturing, food prep, and material moving. Two automation pressures, opposite ends of the education distribution, same country.
Type a job title or SOC code. The card shows what BLS projects, what each exposure measure says, and where the measures disagree. The horizontal bars are the editorial signature: you see the disagreement on your own job in the first ten seconds.
AI exposure on the x-axis, robotics exposure on the y. Bubble size is 2024 employment. Color is BLS 2024–34 projected change in employment. The chart that answers the headline question. Click any bubble to drill into its card.
Three peer-reviewed exposure measures rank the same occupations significantly differently. Eloundou-vs-Felten correlate r ≈ 0.79. Eloundou-vs-Webb (excluded as an outlier in Yale's PCA) correlates only ~0.3–0.5. The disagreement is largest precisely for the high-exposure occupations the dashboard most cares about.
Eloundou et al. (2023, OpenAI) rates each O*NET task on whether a GPT-4-class model could perform it faster with a 50% reduction. Beta scores include software complements. Tends to rate cognitive analysis highly.
Felten, Raj & Seamans AIOE (2023 update) maps ten machine-learning capability dimensions onto O*NET ability requirements. Captures perception and language modeling. Less sensitive to reasoning chains.
Yale Budget Lab (Feb 2026) harmonizes seven leading measures via PCA. Surfaces the latent "AI exposure" axis when researchers disagree on weights. Excludes Webb as an outlier.
What we have measured, as distinct from what we have modeled. Productivity studies show large gains in controlled settings. Aggregate displacement has not yet appeared in US labor statistics. The strongest measured signal is in entry-level employment for the most exposed occupations.
BLS quietly revised exposed occupations down
Same occupations, two projection vintages. Paralegals, customer service reps, programmers, translators, bookkeeping. BLS Aug 2025.
Entry-level cognitive work down 13%
Workers 22–25 in highest-AI-exposure occupations vs older workers in same occupations. Brynjolfsson, Chandar & Chen. ADP payroll, late 2022 – mid 2025.
Firm AI adoption climbing, not a cliff
Census Business Trends and Outlook Survey. Production-use wording: 3.7% (Sep 2023) → 10% (Sep 2025). Broadened wording: 17.3% (Nov 2025).
Real, large, but in lab conditions
| Study | Domain | Effect |
|---|---|---|
| Brynjolfsson · Li · Raymond | Customer support | +14% |
| Peng et al. | Coding (controlled) | +56% |
| Noy & Zhang (Science) | Writing | +40% |
| Cui et al. (MgmtSci) | Field experiments | +26% |
| Dell'Acqua et al. (BCG) | Consulting tasks | +25% |
Treatment vs. control. The controlled-setting result has not translated into aggregate displacement at population scale.
Pick any two occupations. The seven featured pairs each illustrate a different mechanism. The first one is the cleanest illustration of the inversion.
notebooks/jobs_lab.py ↗ joins three peer-reviewed exposure datasets on six-digit SOC, applies a hand-curated BLS layer (2024 employment, May 2024 median wage, 2024–34 projected change, education tier) for the ~85 occupations covering most of US employment, and writes occupations.json. The robotics score in v1 is a heuristic by SOC major group. Webb's pct_robots requires direct BLS access we cannot complete in this build session.
AI exposure is a composite z-score of three peer-reviewed measures: Eloundou et al. β (with software complements), Felten AIOE (capability mapping), and Yale Budget Lab PCA (harmonized across seven measures). Robotics exposure is a major-group heuristic, not a measured per-occupation score. BLS employment, wage, and projected change are the published values from the August 2025 release.
Eloundou et al. · GPTs are GPTs ↗
Felten · AIOE data appendix ↗
Yale Budget Lab · AI Exposure: What do we know ↗
Brynjolfsson, Chandar & Chen · Canaries in the Coal Mine ↗
BLS Employment Projections 2024-34 ↗
Exposure is not displacement. Tasks are not jobs. Most exposure scores predate frontier multimodal and agentic capabilities. Robotics adoption is slower than software adoption. Augmentation may dominate displacement for incumbents while juniors face compressed hiring. New BLS data lands annually; this lab uses the August 2025 release.
How to read these scores ↓
Exposure. A ceiling on what fraction of an occupation's tasks an AI system could perform with current capability. Says nothing about whether employers will substitute capital for labor.
Displacement. The post-adoption employer decision. Mediated by cost, regulation, organizational inertia, and customer acceptance. Often much smaller than exposure.
Augmentation. Productivity gains for remaining workers. Brynjolfsson et al. found +14% on customer support. The largest empirical effect is on novices closing the gap with top performers.
The current empirical state (April 2026). Productivity studies show large gains in controlled settings. Aggregate employment in exposed occupations has not collapsed. The strongest measured signal is a ~13% relative employment decline for ages 22–25 in highest-exposure occupations since late 2022 (Stanford / ADP).