Methodology and sources

What does AI exposure measure?

The ILO's 2025 index (Gmyrek et al., WP 140) scored every one of the 3,265 tasks of the 427 occupations in the international classification (ISCO-08) separately: between 0 and 1, how well a current generative AI model could perform it. Scoring combines several large language models with human expert validation; the occupation's score is the average across tasks — shown here on a 0–100 scale.

The gradient (not exposed, minimal, 1–4) considers the average and the spread across tasks together. In gradient 1 a few tasks are highly exposed; in gradient 4 almost all are. Exposure is NOT a forecast of job losses: it shows where the technical potential for transformation is largest. What happens depends on adoption costs, regulation, employers' decisions and workers.

A task's final score is the average of two AI models' assessments (GPT-4o and Gemini), with expert review for some tasks. On occupation pages the range between the two models' scores is shown next to each task score — that is the uncertainty. The change since the ILO's 2023 index is shown too.

Job titles and search

The English, German and Hungarian names of the 427 ISCO-08 units come from the EU's ESCO system. On top of that, search knows almost 3,000 ESCO occupation names and synonyms in all three languages, each pointing to an ISCO unit. National classifications (e.g. Hungary's FEOR-08, Germany's KldB) map onto ISCO-08.

Employment by country

The 27 EU member states, Switzerland, Norway and Iceland: Eurostat Labour Force Survey (lfsa_egai2d), age 15–74, thousands, ISCO-08 2-digit (43 occupational groups). The United Kingdom has not been in Eurostat since 2019: its data here are ILOSTAT's equivalent breakdown (from the UK Labour Force Survey).

The world map uses the ILO's modelled estimates (ISCO-08 1-digit, ~180 countries) — coarser, but built with the same method for every country; the world total (3.6 billion employed) also comes from it.

National exposure — and why it is an estimate

4-digit (occupation-level) employment by country is not public. So each 2-digit group gets the UNWEIGHTED average of its ISCO units (mean score, share of exposed units, distribution across gradients), and the country value is the employment-weighted average of the groups.

Check: with this method 25.4% of the world's employed work in exposed occupations — the ILO's own, more detailed calculation gives 24%. For high-income countries the ILO reports 34%; we get EU27 32.6%, Germany 34.9%, the UK 34.3%. The estimate gets the order of magnitude and the ranking right; don't take the decimals literally.

Why not finer: employment at 3–4-digit level by country only exists in non-public Labour Force Survey microdata (the ILO uses those too). Weighting with another country's occupational mix (e.g. the US) would add a new, non-European assumption, while the current estimate already matches the ILO's figures. So the interface shows the estimate rounded to a whole percent, marked with “~”.

Unemployment, vacancies, AI adoption

Unemployment rate: Eurostat une_rt_m, seasonally adjusted, monthly; youth: under 25. Job vacancy rate: Eurostat jvs_q_r21 (industry, construction, services), quarterly, not adjusted. Employment rate: lfsi_emp_a, age 20–64. AI adoption: isoc_eb_ai — share of enterprises with 10+ employees using at least one AI technology (no data for Switzerland and the UK).

Countries may refer to different periods (some report later) — the tooltip and the table always show the period. World unemployment is the ILO's modelled estimate.

Wages

The wage block on occupation pages comes from Eurostat's Structure of Earnings Survey (earn_ses22_21): mean gross monthly earnings, enterprises with 10+ employees, all ages. The survey runs every four years; the latest is 2022. Pay by occupation is only published at ISCO-08 1-digit level (9 major groups), so the occupation page shows the pay of the job's broader major group, not of the job itself. Countries are compared at purchasing power parity (PPS). The United Kingdom is not covered.

Personal calculator

The “How much does it affect you?” calculator lists the occupation's ILO tasks. When you mark how much you do each one (no / some / a lot, weighted 0 / 1 / 3), we adjust the occupation's ILO score to your task mix: we add the difference between your weighted and the even task average. If you change nothing, you get exactly the occupation's score. It is a rough personal estimate, not a measurement. Everything is computed in your browser; nothing is sent.

Projections to 2035

Cedefop's 2026 Skills Forecast gives employment paths by occupational group (ISCO 2-digit) to 2035 for the EU, the member states, Switzerland, Norway and Iceland, under its baseline scenario. We show two numbers: expansion (2035 − 2025, may be negative) and replacement demand (jobs left by people retiring or leaving the occupation, 2025–2035). Together they make the expected openings.

The forecast is a model, not a promise: it rests on economic, demographic and technological assumptions. Cedefop does not model the United Kingdom.

The WEF employer survey

The World Economic Forum's Future of Jobs 2025 report is based on answers from more than 1,000 employers (14 million workers, 55 economies). We quote the headline figures (170 million jobs created, 92 million displaced by 2030) from the report text. The WEF gives no exact percentages for the fastest growing and declining jobs, so we show the ranking only. Country profiles rest on small samples — they are indicative.

AI-written text

The weekly snapshot and the “What can you do?” section on occupation pages are written by artificial intelligence, solely from the data shown on the site. Every such text is labelled. AI does not touch the numbers, maps or charts. Details on the AI content page.

Updates

Eurostat, ILOSTAT and IMF data are fetched daily; if a source doesn't respond, the previous data stays. News refreshes every 6 hours, occupation analyses continuously (a few occupations a day), the weekly snapshot every Monday. New releases of the ILO index, Cedefop and WEF data are updated by hand — the date shows in the footer.

Limits — what not to read into the data

Exposure measures the technical feasibility of TASKS, not the economics of adoption, legal barriers or employers' intentions. It doesn't see differences within an occupation (a junior vs. a senior accountant). The national estimate works with group averages. Projections are scenarios. For career decisions these are starting points, not answers.