How much evidence is published?
The economic questions are local. A common AI model reaches economies with different occupations, institutions, languages and patterns of production. Understanding that diffusion requires evidence that can be read alongside the statistics used in each country.
Claude Occupation Evidence studies how occupational activities are represented in the public Anthropic Economic Index. It connects published use to a global geography inventory, audited occupational classifications and official employment context. Its central question is how much of an apparent occupational pattern comes from the activity recorded, the information published and the classification used to describe it.
Why this project exists
The Economic Index makes a valuable contribution: it observes activities people bring to an AI system. Many earlier exposure measures begin with what a model might be capable of doing. Observed use provides a different perspective on where capabilities are being taken up. Handa and colleagues established the task-based approach underlying the Index.
Turning those observations into regional evidence requires further work. The source occupation categories are American O*NET categories, even when the conversations originate elsewhere. European employment tables commonly use ISCO; the United Kingdom uses SOC 2020. A translation between these systems contains substantive assumptions about which activities belong together.
Publication also matters. A country that contributes more conversations is likely to have more detailed categories appear in the released data. Treating the number of visible tasks as an adoption ranking can therefore confuse how much is published with how broadly a workforce uses AI.
Three findings that shape the analysis
Published breadth closely tracks platform volume. Across 121 May 2026 geographies, the rank correlation between positive published task counts and the country's share of global consumer usage is 0.996. The corresponding log-log regression has a slope of 1.034 and an R-squared of 0.935. This strong association makes task counts unsuitable as a standalone measure of national adoption. It does not identify the separate effects of sample size, disclosure rules and genuine task composition.
The occupation facet retains substantially more information. Median published mass in May is 79.17% of country consumer usage for occupation categories, compared with 42.25% for task categories. In the United Kingdom, 437 occupations have a positive published share, while 324 catalogue occupations have any positive published task. This is why direct occupation shares are the explorer's primary measure. The source denominator is retained rather than renormalizing the visible categories to 100%.
Classification and construct choices remain consequential. European mapping alternatives change occupational descriptions. Before any European mapping, the US direct occupation shares correlate 0.628 with Anthropic's March observed-exposure score; the task descriptors correlate 0.581–0.587. These are different constructs and observation windows. Neither a crosswalk nor a stronger correlation turns a conversation share into a share of working time automated.
These findings are computed from the frozen evidence tables. The full paper develops their interpretation, and validation explains which comparisons are informative and which reconstruction checks remain unpassed.
The regional contribution
The project preserves the source classification so that users can inspect country differences without an additional mapping assumption. It then makes the European and UK bridges visible: source and destination codes, relation types, alternative routes, common support and official employment populations.
This allows a regional researcher to distinguish three questions. Which occupational activities are represented in published Claude use in a country? How does that description change when expressed in a local statistical classification? What additional evidence would be required to connect it to workers, firms or labour-market outcomes?
Every geography in the registry remains discoverable. Availability, retained publication mass, employment vintage and classification detail are reported separately. A sparse country is not assigned zero AI use, and a country with strong native occupation evidence does not disappear because a compatible employment denominator is unavailable.
What the dataset contributes
The research output is a documented measurement resource: original country and occupation fields; a fixed task catalogue; alternative occupational mappings; official employment context; and independent comparisons with exposure measures and survey activity profiles. The website makes those components usable through country selection, searchable observations, stable links, dataset previews and downloads.
The contribution is complementary to Anthropic's official Economic Index. It concentrates on international interpretation, reproducibility and regional statistical use. It does not claim to be the first international platform dataset or an exact European reconstruction of Anthropic's exposure measure.
From measurement to impact
The current results describe classified consumer conversations. They do not reveal the user's actual occupation, representative workforce adoption, hours saved or a causal effect on wages and hiring.
A subsequent research programme could connect a fixed, audited usage measure to independently measured employment, pay and vacancy outcomes; compare observed use with capability measures; and study changing task requirements in job advertisements. Such work needs historical outcomes, explicit timing, classification bridges, pre-trend analysis and a credible account of competing economic changes. Those are proposed extensions, not findings of this release.
For now, the practical output is a way to ask better regional questions and inspect the evidence supporting them. Explore a country, read the methods, or choose a dataset.