How the atlas is built

Every number on this site is either sourced (published by a statistical agency or research organisation) or modelled (estimated by combining sourced data with documented assumptions). Modelled figures always carry a MODEL badge; sourced figures carry their source’s name.

Data sources

Data last rebuilt 22 October 2026. Boundaries follow Natural Earth’s de facto conventions and imply no position on disputed territory.

1. Country exposure

We follow the approach of the ILO’s 2025 refined index of occupational exposure. Exposure is a property of occupations (their task content), so a country’s exposure depends on how its jobs are spread across occupations.

• For each ISCO-08 major group we start from the share of its detailed occupations that the ILO places in one of its four exposure gradients (for example 28 of 29 clerical occupations and 39 of 92 professional occupations).
• We apply those rates to each country’s real 2025 occupational mix from ILO modelled estimates, separately for women and men.
• Two multipliers (cognitive groups 1–4; all others) are fitted so the aggregates match the ILO’s published shares. The fit reproduces 35% for high-income countries (ILO: 34%) and 12.8% for low-income countries (ILO: 11%). The world total comes out at 21.2%, against the ILO’s “one in four” (about 24%). The gap arises because the ILO works with 4-digit occupations, which we cannot see in the 1-digit data.
• As an independent check (not fitted), the model gives 42.6% of women’s jobs and 28.7% of men’s jobs exposed in high-income countries; the ILO reports 41% and 28%.
• Highest-exposure jobs are the share of jobs in the ILO’s top gradient. Most of these are clerical. A single scale factor sets the world total to 3.3%. For this band the model gives women 4.1% and men 2.8%; the ILO reports 4.7% and 2.4%.

2. Sector exposure

The ILO publishes employment by sector and by occupation, but not the two combined for every country, and no source publishes GenAI exposure by sector worldwide. We give each sector a typical occupational make-up, broadly in line with industry–occupation tables from the EU Labour Force Survey and US OEWS. Finance, for example, is mostly clerical, professional and technical work; agriculture is mostly farm and elementary work. For each country this sector × occupation matrix is then rescaled by iterative proportional fitting, so its rows match the country’s real sector totals and its columns match its real occupation totals. A sector’s exposure follows from the occupations inside it, and sector figures add up exactly to the national total.

ICT & software as its own sector. The ILO’s modelled estimates publish transport & storage (ISIC H) and information & communication (ISIC J) only as one combined total. We split it using each country’s latest labour force survey ratio of J to H + J (158 countries, mostly 2022–2025 surveys). The 25 countries without one use the median ratio for their income group (LIC 8%, LMC 11%, UMC 20%, HIC 40%). Section J covers software and IT services, telecoms, publishing and broadcasting; the ILO does not publish software alone for every country.

3. Regions

No source publishes sector employment or AI exposure for every first-level region in the world, so regional exposure is modelled. AI usage, by contrast, is measured where available.

• Population comes from Wikidata for 3,843 of 4,579 regions (the most recent preferred statement). For the rest, the remainder of the World Bank national total is spread by area; these figures are marked “≈”.
• Settlement pattern. Each region gets an urbanity index. It is weighted 75% on the share of its people living in its major cities (Natural Earth metro populations) and 25% on population density, each as a z-score within the country, with a bonus for the capital region. Density alone would rank densely farmed regions as urban.
• Sector mix. The national sector mix is tilted by the density index. Denser regions get more finance, ICT and professional services; rural regions get more agriculture (the “urban tilt” column above). The result is then rescaled by iterative proportional fitting, so regional jobs add up to each region’s population share of national employment and each sector adds up to its real national total.
• Exposure. Regional sector jobs × the country’s sector exposure rates. Regional figures therefore always sum to the national total.
• AI usage comes from the Anthropic Economic Index, which publishes usage for 1,157 map regions, matched on ISO 3166-2 codes. Where usage is published only for a larger unit (for example a French region rather than its départements), every map region inside it shows the larger unit’s figures and is labelled accordingly. Usage intensity is the region’s share of national usage divided by its share of national population. For US states we show Anthropic’s own per-capita index.

What these numbers are not

• Exposure is not job loss. An exposed job is one where generative AI could perform a meaningful share of tasks. Depending on adoption, regulation and management choices, that can mean augmentation, transformation or automation.
• Regional estimates are indicative. They capture the broad urban–rural pattern, not local specialisms such as a mining province or a financial hub outside the capital.
• Usage data comes from one AI product. The Anthropic Economic Index reflects Claude.ai conversations, which skew towards particular users and countries. It is a useful signal of where and how AI is used, not a census of all AI use.
• Coverage gaps: the ILO publishes no 2025 modelled estimates for some countries affected by conflict (for example Ukraine and Sudan), so those countries show only sourced indicators.

Coverage

• 243 countries and territories mapped, 183 with exposure estimates
• 174 with the IMF AI Preparedness Index; 121 with Anthropic usage data
• 4,579 first-level regions; 3,843 with Wikidata population; 1,157 with AI usage data

Reproducing the data

The data pipeline is open: npm run data:fetch downloads every source, and npm run data:build rebuilds all files the site uses. The model’s assumptions live in a single configuration file (scripts/model-config.mjs).