Methods · Claude Occupation Evidence

Each stage answers a different question: where platform activity originates, which occupational activities it resembles, how much detail is published and how those categories relate to regional statistics. The methods keep those stages separate.

Observation and source vintage

The primary source is the Anthropic Economic Index release of 26 June 2026, pinned to revision 2ea58ff75e4247d26810c37f10c179edc2466cac. Its April and May 2026 country consumer observations supply the direct occupation and overall-usage fields. The task catalogue is O*NET 30.2. Source release and documentation.

A conversation is assigned an occupational activity by the publisher's classifier. That assignment concerns the activity in the conversation; the user's actual job is unknown. Country evidence concerns the documented consumer surface. The first-party API file has no country dimension and is not allocated to countries here.

Historical task observations use one-week windows in August 2025, November 2025 and February 2026. The April and May observations use calendar months. Sampling and classification also changed. These windows are retained as distinct observations rather than presented as a continuous adoption trend.

Country volume and relative use per capita

Let $U_{cm}$ be country $c$ 's published percentage share of global consumer usage in month $m$ . Let $P_c$ denote the population aged 15–64 in the publisher's population reference. Conceptually, the per-capita index compares country usage share with that country's population share:

A_{cm}=\frac{U_{cm}/100}{P_c/\sum_k P_k}.

The website uses the publisher's released index directly. A value above one means greater platform use per working-age resident than the reference average. It is not the fraction of residents who use AI. Source values retain their published precision.

Direct occupation shares

The primary measure selects the detailed soc_occupation facet, hierarchy level 0 and metric pct. For a published source row with occupation $o$ :

S_{cmo}=\frac{\operatorname{pct}_{cmo}}{100}.

Its denominator is the country's consumer usage in that observation window. Country selection changes the source geography; it does not turn an O*NET category into a national classification.

Let $\mathcal P_{cm}$ be the set of published occupation rows. Retained occupation mass is:

M_{cm}=\sum_{o\in\mathcal P_{cm}} S_{cmo}.

The sum is calculated before any mapping expansion. It is not rescaled to one. The remaining mass can reflect unpublished or unclassified activity and rounding; it cannot be assigned entirely to a single cause from these public tables.

Source status	Meaning	Display and calculation
Published positive	A source row has a positive rounded share	Retain the published value
Published rounded zero	A source row exists and displays zero	Show rounded zero; do not equate it with an absent row
Not published	The catalogue occupation has no source row in that supported country/month	Raw share remains missing; actual usage is unknown

For retained-mass bookkeeping, an unpublished row contributes zero published mass. This does not impute zero actual use. A top-occupation chart shows the original shares; other published occupations and the unresolved residual have different meanings.

The two-decimal source rounding permits an arithmetic range of up to 0.005 percentage points per numeric row when assessing totals. These ranges describe rounding possibilities, not sampling confidence intervals.

Secondary task measures

Let $T_o$ be the fixed set of catalogue tasks associated with occupation $o$ , and let $s_{cmt}$ be a task's positive published share, in fractions. Catalogue task coverage is:

C_{cmo}=\frac{\sum_{t\in T_o}\mathbf 1\{t\text{ has a positive published row in }(c,m)\}}{|T_o|}.

The companion task-share intensity is:

I_{cmo}=\frac{\sum_{t\in T_o\cap\mathcal P^{\mathrm{task}}_{cm}}s_{cmt}}{|T_o|}.

Coverage describes published breadth within a catalogue. Intensity divides released task mass by catalogue size. A task can belong to several occupations, so neither measure can be added across occupations to recover a national usage total. Neither denominator is employment or working time.

European mappings: semantic means

The typed ESCO–O*NET crosswalk contains exact, narrow and broad relations. Alternative routes are kept separate. The source library contains 4,253 typed links, including 498 exact links. Counts of links describe the library; they do not measure the share of employment covered by an exact correspondence. European Commission crosswalk.

For mapping route $a$ , the current ESCO descriptor is the equal mean of available O*NET donor values. If $D_e^{(a)}$ is the donor set for ESCO occupation $e$ :

\overline S_{cme}^{(a)}=\frac{1}{|D_e^{(a)}|}\sum_{o\in D_e^{(a)}}S^{\mathrm{published}}_{cmo}.

ISCO4 descriptors are equal means of represented ESCO occupation descriptors. Coarser ISCO groups use direct means over represented ISCO4 groups. These are taxonomy means with explicit mapping support. An unavailable route remains missing.

They are nonadditive descriptors. Their sums are not national conversation shares, and they cannot be divided by European employment shares to produce a valid usage-to-employment ratio. Alternative routes are sensitivity scenarios, not components of one total or statistical confidence intervals.

Official employment and the native US ratio

Employment context preserves each source's population, classification, year and quality flags. Eurostat and ILOSTAT often provide ISCO groups; UK APS uses SOC 2020. The ONS coding-index route involves lexical weights that should not be interpreted as observed worker transitions between classifications.

For the US, O*NET children are aggregated to their native six-digit SOC code: direct occupation shares are summed, while task descriptors are averaged. Source shares are summed at their original integer rounding precision before conversion to fractions. Exact employment code matching avoids a European bridge.

The BLS National Employment Matrix supplies 2025 base-year employment, including unincorporated self-employment. It counts jobs, not individual workers. Its official total is 170,280,800 jobs; 772 matched detailed groups cover 92.53% of that total. The 2035 projections are not employment outcomes in this analysis. BLS definitions.

For a supported native group with employment $E_o$ and official national total $E_{\mathrm{national}}$ , the descriptive concentration ratio is:

R_{cmo}=\frac{S_{cmo}}{E_o/E_{\mathrm{national}}}.

A ratio above one means the group's share of classified consumer use exceeds its share of national jobs. It does not estimate a worker's probability of using AI. The national denominator is retained; unmatched employment is reported separately.

A prospective allocation method

A future European share construction would require source-to-target weights that preserve mass, including an unmatched destination:

\widetilde S^{(a)}_{cmg}=\sum_o w^{(a)}_{og}S_{cmo},\qquad \sum_{g\,\mathrm{including\ unmatched}}w^{(a)}_{og}=1.

This equation describes a research proposal, not a published result of the current dataset. Weights require a defensible interpretation, independent review and sensitivity analysis. Employment-based weights can mechanically affect a later usage-to-employment ratio; simply forcing a table to sum to one does not validate its allocation.

Reproducibility and interpretation

Frozen inputs, unique observation keys, source precision and publication states are retained in the data outputs. Comparison tables report common occupational samples. Source attribution and redistribution rules remain attached to downloaded material.

The independent reconstruction of Anthropic's exposure formula remains below its prespecified acceptance threshold. Validation separates that result from the successful construction of direct published shares and from partial survey-content comparisons. The data catalogue provides the underlying files and field definitions.