How much evidence is published?
Generative-AI systems reach economies with different occupational structures, institutions, languages and classifications. Public interaction records provide evidence about activities brought to a platform, but interpreting those records as occupational statistics requires careful attention to publication, denominators and mapping.
AI Use and Occupations studies those measurement choices across 121 geographies with May 2026 occupation evidence. Europe is the main classification case because O*NET-based source categories must be translated into ESCO, ISCO and UK SOC categories. The United States is the native benchmark because source occupations can be joined directly to SOC employment codes. This design places the European audit within a global evidence inventory rather than treating Europe as the universe of observed use.
Research question
How much of an occupational pattern in published AI activity reflects the recorded activity, the information released and the classification used to describe it? The analysis separates country usage volume, relative per-capita platform use, direct occupation shares and fixed-catalogue task measures. It then examines how crosswalks, employment joins and external comparators change their interpretation.
Three findings
Published breadth closely tracks platform volume. Across 121 May 2026 geographies, the Spearman correlation between positive published task counts and global consumer usage share is 0.996. A log-log regression gives a slope of 1.034 and . Task-cell counts therefore combine substantive breadth with the publication process and should not be read alone as a ranking of workforce adoption.
The occupation facet retains more published mass. Median retained mass in May is 79.17% for occupation categories and 42.25% for task categories. Direct occupation shares are therefore the primary descriptor, while catalogue task coverage is a secondary measure of published breadth. The source denominator is retained; visible categories are not renormalized to 100%.
Classification and construct choices matter. European mapping routes change occupational descriptions. Before any European crosswalk, US direct occupation shares correlate 0.6284 with Anthropic's March observed-exposure score; the two task descriptors correlate 0.5814–0.5869. These measures differ in window, population and construction, so the correlations are evidence about construct alignment rather than equivalence.
The United Kingdom, Germany and Türkiye share the same two leading published activity labels in May, and the United Kingdom and Germany share 12 of their top 15. This descriptive stability could reflect common platform use, source-classification behaviour or both. Distinguishing those explanations requires multilingual and local-task validation.
What is measured
The observations are country-tagged consumer conversations assigned to occupational activities by the publisher's classifier. The dataset records the source share, whether a row was published, retained mass, task-catalogue breadth and mapping support. It also links compatible classifications to official employment context without converting mapped donor means into national usage shares.
The 250-entry geography registry contains 248 UNSD M49 countries and areas plus Kosovo and Taiwan. Of these entries, 180 have at least one source row, 128 appear in historical task observations, 121 have May occupation evidence and 114 have April occupation evidence. Official employment data are available for 182 entries overall; the intersection with occupation evidence is 113 in May and 106 in April.
These counts describe different, partly overlapping universes and should not be collapsed into one coverage number. The Global availability and Geography support preserve the definitions.
What is not measured
The data do not reveal a user's occupation, the fraction of workers using AI, hours automated, productivity gains or causal changes in hiring, pay or employment. Unpublished rows are unknown rather than zero. Reporting grades summarize observable publication and employment support under explicit post-review rules; they do not validate national representativeness.
The paper develops the evidence and literature, the methods define each quantity, and the technical data note documents construction and files. Terms used across the project are collected in the glossary.