Technical data and construction note

Fatih Kansoy
9 September 2026

Abstract

This note documents the construction, keys, denominators, classifications, missingness rules and public files for AI Use and Occupations. The release combines pinned country-level platform records, a fixed ONET task catalogue, alternative ONET–ESCO–ISCO and UK SOC routes, and official employment context. It contains 121 countries and areas with direct occupation observations in May 2026 and 114 in April. Direct source shares remain distinct from semantic donor means and native employment joins. The note is a file-level reference for reproduction and reuse; substantive results and their relation to the literature are developed in the companion paper.

Keywords: data documentation; occupational classification; O*NET; ESCO; ISCO; SOC; reproducibility; missing data

1. Release identity and scope

The research release is 1.1.0. Its primary platform source is the Anthropic Economic Index revision 2ea58ff75e4247d26810c37f10c179edc2466cac, released on 26 June 2026. The source commit, release directories, retrieval times and hashes are recorded in the project manifest. The fixed occupation task catalogue is O*NET 30.2.

The public catalogue contains 27 curated evidence tables. Each entry records the source path, public download, format, row count, primary key, grain, units, schema, file size and SHA-256 hash. A catalogue preview is not a sample file: each linked download contains the complete frozen public table.

The observations concern classified consumer platform activity. The source assigns conversations to occupational activities; it does not observe the user's job. The first-party API source has no country breakdown and is not allocated to countries in this release.

2. Observation windows

The archive contains five country observation windows drawn from four selected releases.

The June source documents the monthly end dates as exclusive. Earlier files retain the publisher's displayed weekly date fields. Sampling, classifiers and taxonomy differ across releases, so the construction does not fill missing months or treat the five windows as a uniform panel.

The Source country period inventory table contains the source availability by geography and period. The Country coverage table records derived task support for the same 581 geography-period observations.

3. Geography registry and coverage universes

The registry has 250 entries: 248 countries and areas extracted from UNSD M49, plus Kosovo and Taiwan added as explicit source geographies. Geography keys use ISO alpha-3 style identifiers where available. Publisher labels that are not geographic, including NONE and not_classified, do not become registry entries.

Coverage counts refer to distinct universes:

The May roster includes 118 UN member states, the State of Palestine, Puerto Rico and Taiwan. The categories describe the statistical source and are not used to infer political status.

The Country registry table is the geography spine. The Global availability table summarizes source and derived coverage. The Labour availability table records labour-statistics availability separately.

4. Source selection and row keys

The primary country values select the consumer surface and retain the publisher's date, geography, facet, hierarchy, metric and category fields. Direct occupation evidence uses the soc_occupation facet at hierarchy level 0 with metric pct. Source labels, codes and displayed precision are retained before any mapping.

At source grain, a row is identified by the available combination of release, surface, observation dates, geography, facet, hierarchy, metric and category. Derived tables use narrower documented keys. Key checks reject duplicates rather than choosing an arbitrary row.

The core public tables have these grains:

The O*NET, ESCO and ISCO occupation panels above contain monthly observations. The native US panel also retains earlier weekly task descriptors; its period field is the BLS employment year, while date_start identifies the platform observation window. Measures are columns in these wide tables, not components of their row keys. The data catalogue is authoritative for the exact primary key of each released table.

5. Missingness and source precision

Three publication states remain distinguishable:

An absent category contributes zero to the sum of published mass but is not assigned zero actual use. The residual outside the published rows may combine unpublished activity, unclassified activity and rounding. The public source does not identify their separate contributions.

Source percentages are released to two decimal places. A reported value has a row-level rounding half-width of at most 0.005 percentage points. Summed rounding ranges accumulate these arithmetic bounds over numeric rows. They are not sampling intervals and do not account for classification or disclosure uncertainty.

6. Direct occupation measures

For geography $c$, month $m$ and published occupation $o$, the direct share is

The denominator is total consumer usage for the geography in that observation window. Retained occupation mass is

where $\mathcal{P}_{cm}$ is the set of published occupation rows. The source shares are summed before mapping and are never renormalized to the released categories.

The Direct occupation source download preserves only published source rows. The larger O*NET occupation panel expands to the fixed occupation universe and carries publication-state fields so that absent rows can be distinguished from rounded zeros.

7. Task catalogue measures and historical bridge

For fixed task set $T_o$, let $s_{cmt}$ be the source task percentage divided by 100 and let $\mathcal{P}^{\mathrm{task}}_{cm}$ be the set of published task rows. Task coverage is

and task-share intensity is

All catalogue tasks remain in the denominator. A task may be listed under several occupations, so these measures are not additive across occupations.

Historical task text first follows a legacy-ID bridge into the current catalogue. Only unmatched source cells may use exact normalized current-task text as a fallback. The procedure recovers 1,839 source cells, with zero observed convergent collisions in the frozen audit. Strictly matched rows are not expanded using a union of every text candidate. Changed text under a retained ID and identical text attached to multiple occupation-specific tasks remain flagged.

Source-share floors of 0.01, 0.05 and 0.10 percentage points are supplementary sensitivity checks. They remove small positive cells while holding the catalogue denominator and mapping route fixed. They do not reconstruct the publisher's count threshold.

8. European mapping construction

The typed O*NET–ESCO library contains 4,253 relations, including 498 exact links. Routes retain exact, exact-plus-narrow and all-typed relations separately; an 8,627-link untyped alternative supplies a sensitivity route.

For route $a$ and ESCO occupation $e$, with available donor set $D_e^{(a)}$, the descriptor is

Here $S^{\mathrm{published}}_{cmo}$ equals the source share when a row exists and zero published mass for an absent row within a supported geography-month. The donor set includes all linked O*NET catalogue occupations with a defined published-mass value, including donors with no published row. Donors outside the available catalogue are excluded. This convention measures published mass; actual use for an unpublished donor remains unknown.

The same donor-mean operation is applied to task descriptors. A missing donor set remains missing. ISCO4 values are equal means across represented ESCO occupations. ISCO1 and ISCO2 values are direct means over represented ISCO4 groups, avoiding a nested average that would assign implicit extra weight to small branches.

Mapped values are semantic donor means and are labelled as such. They do not conserve source mass and cannot be interpreted as allocated national shares. Alternative routes describe mapping sensitivity rather than sampling uncertainty. The ESCO occupation panel and ISCO occupation panel files retain route, donor count and support fields.

9. Employment joins and conditional summaries

For ISCO sources, the pipeline selects the finest eligible detail while retaining the actual year, sex and age scope, survey population, ICLS definition and publication flags. Finer observations can be older than an available coarse series. No source is silently relabelled to a common population or year.

For an available mapped score $x_g$ and employment $E_g$, a conditional weighted mean is

The matched-employment fraction is

Different routes may have different matched denominators. Weighting a donor mean does not convert it into a national conversation-share allocation.

The United Kingdom uses APS employment and an open ONS SOC 2020 coding-index route. Counts of coding-index job titles supply lexical weights; they are not observed transition probabilities between SOC and ISCO. Earnings sensitivities use Eurostat SES 2022 and UK ASHE only where compatible values are available. Combining earnings and employment from different source populations does not produce a measured wage bill.

10. Native US construction

O*NET children are aggregated to their native six-digit SOC code. Direct occupation shares are summed at source integer precision, while task coverage and intensity are averaged within SOC code. The BLS National Employment Matrix then joins on exact detailed codes.

The Matrix contains 831 line items and an official 2025 total of 170,280,800 jobs. Exact matching supports 772 detailed groups and 92.53% of the total. Aggregate and unmatched BLS codes are not split through a synthetic concordance. The Matrix population counts jobs, including unincorporated self-employment.

For matching occupation $o$, relative concentration is

The national denominator retains unmatched employment. The ratio is a descriptive comparison of shares, not a probability of worker use. Construction inputs and outputs are in US bls employment, US native soc panel and US native weighted summary.

11. Exposure proxy and acceptance gate

For task $t$, let $C_t^{\mathrm{work}}$ be work-related consumer counts, $A_t$ global API counts, $q_t$ consumer automation share and $\beta_t$ capability. The published construction begins with

and, when $W_t>0$,

The gated task value is

with $r_t=0$ when $W_t=0$, without evaluating the zero denominator. For occupation-specific task-time weights $w_{ot}$, exposure is

Both sums range over the occupation's full task set. Tasks that fail either gate contribute zero to the numerator but remain in the time-weight denominator; weights are not renormalized over eligible tasks.

The frozen independent proxy reaches Spearman 0.89465 on 756 occupations with 6 common top-ten occupations. The prespecified numerical gate requires correlation at least 0.95 and at least 8 common top-ten occupations, so the proxy fails. Equal aggregation of already-published task scores reaches 0.87154 but is a downstream diagnostic, since those inputs embed publisher outcomes.

Exact reconstruction is also blocked by unavailable original time fractions, work-share imputation, similar-task grouping, employment allocation and uncensored intermediates. The failed local-model time weights are excluded from substantive results. Their preservation in the audit archive permits deterministic checking of why they failed; it does not turn them into valid estimates.

12. Validation and comparison tables

Comparison outputs retain pairwise common support and the direction of each source measure. They do not flip signs or select coefficients according to favourability.

PIAAC public activity cells require at least 30 valid respondents. Survey weights enter activity shares, while the released rank correlations use occupations as their comparison units. Restricted respondent records are not included.

13. Reporting grades

Grades combine observable publication and employment components: positive task-cell breadth, retained occupation mass, two monthly occupation observations, employment year, classification detail and matched-employment fraction. The rules yield 29 grade A, 11 grade B, 81 grade C and 129 grade D registry entries.

The grades were adopted after inspecting the public data. Their cutoffs are practical reporting thresholds rather than preregistered sample rules. A grade does not estimate representativeness, classifier accuracy or latent data quality. The exact components are released in Geography support and should be used instead of the letter alone when selecting a sample.

14. Integrity checks and provenance

The build checks unique keys, finite values, permissible ranges, publication states, aggregation denominators and source-to-public row counts. Numerical table generation, source licensing, PDF layout and scientific validation are separate checks with separate receipts.

File hashes identify the precise public bytes. Source manifests identify upstream URLs, commits and retrieval times. A hash establishes identity, not scientific validity. When source-derived files are updated, the catalogue, data dictionary and validation receipts must be regenerated together.

Restricted mapping vectors, comparator vectors and respondent-level PIAAC records remain outside the public package. Public aggregate files carry component-specific attribution and redistribution terms. The catalogue licence field links those terms rather than asserting a single licence for the combined archive.

15. Recommended entry points

Readers reproducing headline results should begin with Facet comparison, Volume regression and Geography support. Classification analysis begins with the O*NET, ESCO and ISCO panels. Native US work begins with the US SOC panel and BLS employment file. Validation work should use both the summary file and the underlying pair-level or geography-level observations.

The data catalogue supplies schemas and downloads for all 27 tables. The companion paper presents substantive results and scholarly interpretation. Shared terminology is defined in the glossary, and all cited sources are collected in the bibliography.

References