An interactive research and art project mapping AI's technical capability to replace human labor — from the first point at which global employment can be measured to the furthest point at which it can be reasonably projected.
Created by Lukas Seel and Enya Trenholm-Jensen. Open source. April 2026.
The Large Labor Model traces human labor across 15 territories of work (13 modern + 2 historical aggregates) from 1800 to 2041. It maps two things:
Labor shares — how many people work in each territory, globally, based on ILO data and historical reconstructions.
Replaceability — when AI or robotics becomes technically capable of performing each territory's core tasks, scored 0–100 based on what is commercially available at the frontier.
The gap between what AI can do and what has actually changed in the labor market is the project's central provocation. A territory can score 90 while employment barely moves. That is not a flaw — it is the most important thing the visualization shows.
This is a model, not a prediction. Every number is an invitation to engage, refine, and challenge.
Replaceability = can the best commercially available AI/robotics perform this role? Pure technical capability. Lab prototypes don't count. Deployment rates, economics, regulation, and social preference are recorded as barrier annotations — they do not suppress the score. This is what the model measures.
Replacement = when workers actually lose jobs. Lags replaceability by months to decades. This is what labor share data tracks.
| Territory | ISIC | What it covers |
|---|---|---|
| Land & Sea | A, B | Agriculture, forestry, fishing, mining |
| Making Things | C | Manufacturing |
| Building Things | F | Construction |
| Moving Things | H | Transport, logistics, warehousing |
| Buying & Selling | G | Retail and wholesale trade |
| Money & Data | J, K, L | Finance, insurance, information, real estate |
| Care & Health | Q | Human health and social work, personal care |
| Learning & Teaching | P | Education |
| Making Meaning | R | Arts, media, creative industries |
| Governing & Protecting | O | Public administration, defense |
| Feeding & Hosting | I | Accommodation, food service |
| Maintaining & Fixing | D, E, S | Utilities, repair, personal services |
| Thinking & Leading | M, N | Management, consulting, science, professional strategy |
Plus 2 historical aggregate territories (Industry and Services) that cover the period before granular sectoral data was available — they disaggregate progressively into the 13 modern territories between 1900 and 1980.
Two principles shape how occupations are placed into territories:
-
Nature of work, not industry sector. A barber's work is personal care, so barbers belong to Care & Health regardless of whether they work in a salon (retail) or a hotel (hospitality). A car mechanic's work is repair, so mechanics belong to Maintaining & Fixing regardless of whether they work for a manufacturer or a dealership.
-
Functional managers go with their function. A Finance Manager belongs to Money & Data, not Thinking & Leading. Thinking & Leading is reserved for organizational generalists (Chief Executives, Management Consultants), pure knowledge workers (Mathematicians, Sociologists, Humanities Academics), and people functions (HR Specialists, Development Officers).
Current version: 5.0 (April 2026) — a ground-up rebuild with task decomposition, empirically-calibrated replacement formula, and editorially-reviewed moderate scenario. See docs/v5_changelog.md, docs/v5_methodology.md, and docs/v5_editorial_process.md. v3.2 remains in archive/ for lineage.
| Layer | Records | Coverage |
|---|---|---|
| Labor shares | 635 | 1800–2041 |
| Replaceability scores (territory-level) | 936 | 1970–2041 (585 historical 1970–2014 + 351 modern 2015–2041) |
| Occupations (per-occupation scores) | 480 | ISCO-08 4-digit + 9 specialized splits, searchable |
| Tasks (per-occupation task decomposition) | 4,818 | 2–14 tasks per occupation, 6-vector classified with difficulty + time weight |
| Displayed occupations (canvas) | 104 | 8 per modern territory |
| Technology events | 77 | 1764–2026 |
| Historical occupations | 24 | 8 per historical aggregate (3 categories) |
| Sources | 6,687 | Per-record citation |
Source architecture:
- ILOSTAT modeled estimates (1991–2025, primary modern labor)
- World Bank / IISS Military Balance (1988–2025, Governing & Protecting)
- GGDC 10-Sector Database (1870–1991, splice-adjusted to ILOSTAT classification at 1991)
- Bairoch, Maddison, Mitchell (pre-1870 reconstructions)
- O*NET + BLS OES/OOH + live LinkedIn scrapes (task decompositions + common titles)
- IFR World Robotics 2025 (P_A calibration)
- METR, Epoch AI, SWE-bench, GPQA, HLE, ARC-AGI-3 (benchmark calibration landscape)
- Anthropic Economic Index (Massenkoff & McCrory, March 2026 — cross-section for Phase 6 / 11 calibration)
- Waymo + humanoid pilots (Figure at BMW, Tesla Optimus, Amazon Vulcan — deployment ground truth)
- Every occupation decomposed into tasks, each tagged with a capability vector (one of six: C_R routine cognitive, C_G generative cognitive, P_A physical automation, Phi_S selective physical, Phi_U unstructured physical, S_E system engineering), a difficulty threshold, and a time weight.
- 2026 capability values (C_R=0.76, C_G=0.57, P_A=0.75, Phi_S=0.46, Phi_U=0.15, S_E=0.35) set by two independent recalibration instances under an honest deployment lens (benchmark-saturation ≠ deployment maturity). Tight convergence between Opus 4.7 and GPT-5.4.
- 1970 → 2041 trajectory: two parallel forecasters + reconciler at each end — Phase 7 forward (2026 → 2041), Phase 12 backward (1970 → 2025). Reconciled under
mid = min(A_mid, B_mid)to prevent any single instance's aggressiveness from pulling the central case. - Per-occupation barrier tags: REGULATORY / ECONOMIC / HUMANOID_DEPENDENT / HUMAN_PREFERENCE / NONE. Regulatory and preference barriers live in the replacement formula, not in the replaceability score itself.
The visualization shows a single projection under the moderate capability trajectory. The displacement model converts capability gains into projected workforce reduction using an empirically-calibrated conversion rate (0.30) and a 15-year lag schedule (7/18/30/42/44% absorption) fit against seven historical automation cases and the Anthropic Economic Index cross-section. Low and high scenario bands are preserved per-occupation as structural metadata but are not yet editorially reviewed — v5.0 ships moderate only. See docs/v5_methodology.md for parameters, assumptions, and the full honest-limitations inventory.
The visualization is a single index.html loading a single JSON dataset. No build step, no framework, no dependencies. Canvas-rendered. Deployed on Vercel.
git clone https://github.com/lamentierschweinchen/ai-reach.git
cd ai-reach
python3 -m http.server 8000
# Visit http://localhost:8000ai-reach/
├── index.html # Main visualization (~2,700 lines, canvas + DOM)
├── methodology.html # Short methodology page (linked from the site)
├── methodology-full.html # Full methodology + changelog
├── ai_reach_v5.0.json # Current dataset (the only file index.html fetches)
├── DATA_ARCHITECTURE.md # Schema and rendering pipeline reference
├── LICENSE # MIT (code) + CC BY 4.0 (data)
├── CONTRIBUTING.md # How to propose corrections and additions
├── CITATION.cff # Citation metadata
├── docs/
│ ├── hero-2041.png # Repo hero image
│ ├── v5_changelog.md # Full v5 changelog (primary reference)
│ ├── v5_methodology.md # Full v5 methodology (markdown source for methodology-full.html)
│ └── v5_editorial_process.md # Build journey + Phase 11/12 interventions (post-mortem)
├── archive/ # Historical dataset versions
│ ├── ai_reach_v3.0.json
│ ├── ai_reach_v3.1.json
│ ├── ai_reach_v3.2.json
│ └── README.md
├── fonts/Inter-Variable.ttf
└── favicon.svg
Open source by design. Specific, evidence-backed corrections are exactly what we want. See CONTRIBUTING.md for the full guide.
The short version:
- Score correction: Open an issue using the Score correction template. Include territory, year, current score, proposed score, and a link to the commercially available product or deployment that supports the change.
- Bug report: Use the Bug report template.
- Methodology critique: Open a discussion or an issue. Read
methodology-full.htmlfirst — most of the model's assumptions are documented there with notes on which are most worth challenging.
The standard for evidence is commercial availability, not lab demos. A research paper showing 95% accuracy on a benchmark does not raise a score. A product you can buy — or a documented deployment — does.
Seel, L. & Trenholm-Jensen, E. (2026). Large Labor Model (v5.0) [Dataset].
https://largelabormodel.com
GitHub auto-generates a "Cite this repository" widget from CITATION.cff.
- Code (
index.html,methodology.html,methodology-full.html, all JavaScript): MIT - Data (
ai_reach_v5.0.jsonand successor versions, methodology content): CC BY 4.0 - Inter typeface: SIL Open Font License 1.1
You are free to share, adapt, and build on the dataset for any purpose, commercial or otherwise, with attribution.
Large Labor Model is grounded in early 2026 — a moment when AI capabilities have made extraordinary leaps that most research has not yet caught up with. Helping the model catch up is the work.