The project connects public Anthropic Economic Index summaries to international and UK occupation classifications and official labour statistics. It preserves published usage totals, measures gaps in the evidence and shows how classification choices affect the results.
The observations describe categories of consumer AI activity. They do not identify users’ jobs.
Global publication pattern
Choose a region to show only its countries. Select or hover over a point to see its name and values.
Loading the global publication figure…
How much AI activity is visible?
Published categories cover different amounts of activity in each country. A longer task list can reflect more platform use and more published detail, rather than wider adoption across the economy.
Occupation categories retain more activity
Median published share, May 2026
Country rankings by published task count and usage volume are almost identical: 0.996 rank correlation. Task counts alone are therefore a poor guide to differences in occupational use.
Why connect AI activity to occupations?
Knowing what AI could do is different from observing the activities people bring to it. The Anthropic Economic Index makes some of those activities visible. Connecting them to local occupations and labour statistics makes that evidence more useful for economic research.
But a country label does not make a result locally representative. Job classifications differ, publication is uneven, and an activity label does not identify the person doing it. This project makes those measurement choices inspectable.
- Different job classifications
- Documented links connect US O*NET categories to European ESCO, international ISCO and UK SOC. Allocation preserves the published total and shows the ambiguity in those links.
- Unequal publication across countries
- Coverage tables retain what was published, what is missing and the original denominator. An unpublished category is not treated as evidence of no AI use.
- Activities, workers and outcomes
- Consumer activity stays distinct from employment and pay. Prepared data joins support further research without claiming that activity shares measure worker adoption or economic effects.
From source categories to local occupations.
The method divides each published US-category share among its linked destination occupations. The baseline uses equal splits; stricter links and conditional ranges show how much the choice matters.
Every published share is accounted for
Germany, May 2026. Using all typed links at the ISCO two-digit level:
Equal splits are a transparent assumption, not observed flows of workers. The international allocation is available beyond Europe; Europe and the UK provide detailed cases for connecting it to labour statistics.
One occupation, two populations
ICT professionals in Germany
Information and communications technology professionals · ISCO 25
Both bars use a 0–15% scale.
4.04×Activity share relative to employment share
This group receives a larger allocated share of national consumer activity than its share of employed people (2025 residents aged 15–74). It does not mean its workers are 4.04 times as likely to use AI.
Conditional mapping range: 9.14–13.70% of country activity. This describes alternative splits among permitted links, excludes unpublished activity, and is not a confidence interval.
What the 35 datasets provide.
Download complete tables, preview their contents and read the field definitions and source notes. Coverage differs across datasets: they do not all describe 121 countries and areas, or the same dates and occupations.
Compare activity across countries
Published occupation shares and coverage records show which activities appear, where observations exist and how much of each country’s activity is visible.
Work with international and UK occupations
Allocated shares, every mapping weight and accounting checks let you reproduce the classification step or test another permitted mapping.
Relate activity to labour-market data
Prepared panels connect activity measures to European employment (2020 onward), UK employment and pay (2021–2025), and UK recruitment adverts (January 2017–July 2026). These are research inputs; no hiring or wage effect is estimated.
Compare months and occupation labels
The April–May comparison applies fixed classification weights to common countries and occupation groups, retaining each month’s publication status. An ESCO reference supplies occupation labels in 28 languages for local review and evaluation. Neither establishes a long-run diffusion trend or multilingual classifier accuracy.
Check interpretation against other evidence
US employment benchmarks, theoretical AI-exposure measures and PIAAC work-activity surveys provide distinct checks on what the occupational measures describe.
Tables retain source dates, units, population definitions and quality flags. Reuse means taking those documented inputs into your own comparisons, mapping checks or empirical study.
What is observed, and where?
The project analyses public aggregate summaries of consumer Claude.ai activity. Anthropic assigns activity and occupation categories to conversations; this project does not access private conversation text.
A request such as help me debug this code
describes programming activity. Its author could be a developer, a student or someone working on a personal project. The category does not identify their job or establish workplace use.
The source covers one provider and a selected consumer population. Activity shares are not economy-wide adoption rates, and these data alone do not identify effects on employment, wages or productivity.
The geographic and time coverage
- May 2026
- 121 countries and areas with occupation observations.
- April 2026
- 114 countries and areas with occupation observations.
- Earlier task snapshots
- Three weekly windows, beginning in August 2025, are documented separately. They are not appended to the monthly occupation series as a continuous trend.
- Geographic directory
- 250 entries retain both observed and unobserved places. Directory membership is not evidence availability.
The paper behind the data.
Measuring AI Use across Countries and Occupations
The paper develops the motivation and literature, defines the measures, presents the findings and distinguishes what the evidence supports from what remains unresolved. Five appendices document construction, allocation formulas, additional comparisons, exposure reconstruction and reproducibility.
Questions the evidence opens up
Locally reviewed classification links, clearer publication rules and comparable observation windows can improve regional AI measurement. Independent worker and firm data are needed to study diffusion and economic effects.
The project supplies outcome joins and multilingual reference data for this work. The paper reports the validation evidence, underlying assumptions and remaining measurement limits.