How we measure AI job risk

The short version

• Measured: how many people work in each occupation group, and (where published) wages. These come from official statistics: ILO ILOSTAT, Eurostat, the US Bureau of Labor Statistics, the OECD, and national statistics offices.
• Estimated: the AI exposure, robotics risk, offshoring risk and work-from-home scores (0 to 10). They were written by an AI language model (Claude Opus 4.7, Anthropic) in one scoring session on 28 May 2026, guided by published research. They are informed judgements, not measurements and not forecasts of job losses.
• Derived: a country's AI exposure score is the employment-weighted average of its occupation-group scores. Risk velocity and recovery resilience are indexes built from World Bank indicators (explained below). They are not predictions.

1. Measured data and sources

Employment by occupation uses the ISCO-08 classification. For each country we use the most recent year available in the source, so the year differs by country: 82 countries use 2025 data, 38 use 2024, 13 use 2023, and 72 use 2022 or earlier (the oldest is 2003). The data year is shown on each country page.

• ILO ILOSTAT (CC BY 4.0): employment by occupation for 171 countries.
• Eurostat (labour force survey, table lfsa_egai2d): 36 European countries, plus Structure of Earnings Survey wages for 19.
• US Bureau of Labor Statistics (Occupational Outlook Handbook, May 2025 reference): the United States, with 341 detailed occupations.
• OECD Average Annual Wages (USD, PPP): a wage benchmark for 38 countries. Wage tiers for other countries come from ILO earnings data where it exists; 127 countries have some wage data.
• World Bank open data (198 countries) and UNDP Human Development Report 2025 (177 countries) for economic and human development context.

Most countries are described by 10 broad occupation groups. 34 countries have finer detail (up to 42 groups) and the United States has 341 occupations. Fewer groups means a coarser picture for that country.

2. How the AI scores were made

Each ISCO-08 occupation group received four scores from 0 to 10:

• AI exposure: how much language and generative AI will reshape this work in the next 5 to 10 years (0 = none, 10 = AI can do almost all of it).
• Robotics risk: exposure to physical automation hardware.
• Offshoring risk: how easily the work can be done remotely from another country.
• Work-from-home potential: how much of the work can be done from home.

The scores were produced by Claude Opus 4.7 in a single session on 28 May 2026, informed by published research on automation and AI exposure (including Frey and Osborne, the OECD, McKinsey and the IMF). Scores move in half-point steps. The same group score is applied in every country, so a country's headline score differs from another's because its workers are spread across different occupations, not because the same job is rated differently there.

Country score: sum over occupation groups of (group AI exposure x number of workers in the group), divided by total workers. The result is shown out of 10.

3. How the scores were checked

We compared our scores for the 10 occupation groups that Anthropic published for its US study (Massenkoff and McCrory, March 2026) with that study's figures, using the numbers stored in our data file:

• Our AI exposure score vs the study's theoretical exposure: correlation 0.96 (rank correlation 0.90).
• Our AI exposure score vs the study's observed exposure from actual Claude usage: correlation 0.86 (rank correlation 0.89).

Read this carefully. Only 10 groups were compared. They include very different jobs (software and clerical work against trades and farming), which makes agreement easier to achieve. The study covers the United States only. We have not yet compared our scores with the ILO's task-level global index (Gmyrek and others, 2025); we plan to publish that comparison. Agreement with other measures is evidence that the ranking of broad groups is sensible. It is not proof that any score is correct.

4. Derived indexes: what they are and are not

Risk velocity (0 to 10) is built only from digital infrastructure: fixed broadband subscriptions per 100 people (capped at 50) and secure internet servers (capped at 5,000), combined as a geometric mean. We label ranges of this index as "imminent (1-3 years)", "arriving (3-7 years)", "delayed (7-12 years)" and "distant (12+ years)". These time windows are descriptive labels for ranges of an infrastructure index. They are not forecasts, and they are not based on occupations or on measured AI adoption. Treat them as "how ready is this country's digital infrastructure for AI-driven change", not "when will jobs disappear".

Recovery resilience combines a World Bank human-capital measure with social protection coverage. Demographic alignment and sovereign buffer use age structure, fertility, migration and government debt. They are simple, transparent indexes, not validated predictors.

5. What this cannot tell you

• It cannot say whether a particular job will be lost. Exposure measures how much of an occupation's work AI could affect, not what employers will do.
• Scores are for broad groups. A graphic designer and an architect sit in the same group, so they share a score, even though their work differs.
• Language models can be wrong or inconsistent, and a different model or a later date could produce different scores.
• Data years differ, informal work is under-counted in many countries, and some countries have only 10 broad groups.
• The AI features on the calculator (matching your job title to a group, and the short career insight) are written by an AI model. The score shown always comes from the table above, and the insight text is commentary.

What outside research says about exposure and actual job outcomes

• ILO research brief, February 2026. Exposure indicators "reveal technological susceptibility, not labour market outcomes" and cannot be interpreted as predictions of job displacement. Different exposure indices also vary widely depending on how they are built. (Merola, Ernst, Samaan and others, Workers' exposure to AI, International Labour Organization.)
• Anthropic, March 2026. Massenkoff and McCrory found no systematic increase in unemployment for workers in highly exposed occupations since late 2022, with tentative evidence that hiring of 22 to 25 year olds into those occupations has slowed. Their data covers the United States.
• Stanford Digital Economy Lab, August 2026. Using ADP payroll data through June 2026, Brynjolfsson, Chandar and Chen report that employment of 22 to 25 year olds in AI-exposed occupations is about 19% below where it would be if it had kept pace with less-exposed peers, while the researchers find no widespread, economy-wide displacement. The gap comes mainly through reduced hiring, not layoffs. This is United States data and says nothing directly about other countries.

None of these studies is a forecast, and neither is this site. Exposure scores here describe how much of an occupation's work current AI could affect, which is why we show them next to, not instead of, wages, employment counts and local conditions.

6. Corrections and changes

If you find an error or have evidence that a score is wrong, report it here or email hello@worldjobsdata.com. Changes to scores or method will be recorded here. To cite the dataset, see how to cite.

• 4 October 2026: added the outside research on exposure versus actual job outcomes (ILO, Anthropic, Stanford) and a link to how to cite the dataset. No scores changed.
• 1 October 2026: first published methodology page, including the comparison with Anthropic's published groups and the explanation of risk velocity.
• 28 May 2026: AI scores produced (version 1).