Methodology

The works below are among those behind the figures shown on this site, together with the literature we reference when framing AI exposure. Values are carried at the value each publisher assigned. Presentation is localised — for example, where a publisher expresses a value as a worded category rather than a number, the wording shown is that category translated into the page's language. Localisation changes how a value is written, not which value its publisher assigned. Localisation changes how a value is written, not which value its publisher assigned. Localisation changes how a value is written, not which value its publisher assigned. The Credits & Sources page carries the dataset attributions and licences.

• Massenkoff & McCrory (2026) - Labor market impacts of AIReference
• Eloundou et al. (2023) - GPTs are GPTs: theoretical task exposure framework
• Brynjolfsson et al. (2025) - Canaries in the Coal Mine: employment effects measured from ADP payroll microdata
• U.S. Bureau of Labor Statistics - Employment Projections 2025-2035
• O*NET SOC Classification System - standardized occupation and task taxonomy

The figures defined below are republished from external publishers, and each names its source where it appears. Where an occupation page shows a rank, a share, or a position within a published range, that value states a relation between the publisher's own values rather than a new measurement — AI Changing Work publishes no measurement of its own. This section defines the metrics below as their publishers define them.

OpenAI "GPTs are GPTs" — beta (β)

β = direct exposure (E1) + 0.5 × exposure when tools are available (E2), as defined by the source repository. We show it at occupation level under two rating regimes the authors published — human raters and GPT-4 raters — and at task level as the share of rated tasks scoring β ≥ 0.5. Values are 0–1 as published; we convert to a percentage for display only.

Anthropic Economic Index — observed exposure

A theoretical exposure index weighted by measured Claude usage, per the original report's definition. It describes usage observed at the time the dataset was published. We also show per-task penetration: how many of an occupation's tasks show any observed usage, and their mean. Values are 0–1 as published.

U.S. Bureau of Labor Statistics — Employment Projections 2025–35

Employment, median wage and projected change are transcribed from BLS Employment Projections 2025–35 (Table 1.2), a public-domain federal publication. These figures describe the SOC occupational group, not the individual job title, so occupations sharing a SOC code share them. Employment is published in thousands of jobs.

U.S. Bureau of Labor Statistics — relative AI exposure bands

BLS places each occupation in its National Employment Matrix into one of four relative bands — Low, Moderate, High, Very high — published alongside Employment Projections 2025–35. The band is a rank position among the occupations in that table, not a level, a probability of adoption, or an employment forecast, and it does not separate automation from augmentation. It is also not a first-hand BLS measurement: BLS takes each occupation's percentile ranks from several previously published studies and clusters them into the four bands. Two of those studies are among the datasets this site already carries, so the band and those figures are not independent of one another. We publish the band label as BLS wrote it, translated into the page language; the ordering number BLS does not publish is never shown. The band is assigned per National Employment Matrix (NEM) code rather than per job title, and a NEM code can cover more than one occupation on this site, so occupations sharing a NEM code carry the same band.

AI Changing Work calculates no exposure, risk or automation score of its own. Where a source dataset does not cover an occupation, we show that absence rather than substituting a figure.

Occupations are described at the level of individual tasks, because the published exposure datasets themselves rate tasks rather than whole jobs. This section sets out how those external ratings are attached to the occupations on this site.

Task-Level Decomposition

Each occupation is broken down into constituent tasks so that task-level ratings from the source datasets can be attached to it. The task lists come from O*NET; the ratings attached to them come from the external datasets and are not modified.

External exposure indices

Occupation and task exposure values are reproduced from published research datasets: Anthropic Economic Index (CC BY 4.0), OpenAI "GPTs are GPTs" exposure rubric (MIT License), ILO Working Paper 140 (CC BY 4.0), and BLS AI exposure categories (Public domain (17 U.S.C. §105)). Source values are carried at the value each publisher assigned: AI Changing Work does not rescale, reweight, reclassify or blend them, and calculates no exposure score of its own. Our contribution is the mapping onto O*NET occupation and task codes, and occupations absent from a dataset are reported as absent rather than scored. Presentation is localised — for example, where a publisher expresses a value as a worded category rather than a number, the wording shown is that category translated into the page's language. Localisation changes how a value is written, not which value its publisher assigned.

Employment projections

Employment, wage and projected-change figures are transcribed from the U.S. Bureau of Labor Statistics Employment Projections 2025–35 (Table 1.2), a public-domain federal publication. They describe the SOC occupational group, so occupations sharing a SOC code share these figures. We publish no employment projections of our own.

Transparency about our limitations is essential to responsible analysis. Users should consider these factors when interpreting our data.

No Underlying Sample

Our figures are not derived from a measured sample. There is no survey, telemetry, or panel behind them, so no confidence interval can be stated. Treat every number on this site as an unvalidated estimate.

Geographic & Occupational Coverage

Coverage is limited by the source datasets: of 1,010 published occupations, some are absent from one or more of the exposure datasets and are shown as not covered rather than scored. The underlying data primarily reflects the U.S. labour market and English-language AI interactions, which may not represent global patterns.

Update Frequency

Metrics are revised when we re-run our own analysis. We do not receive automatic updates from any external data provider.

Theoretical vs. Observed Gap

Exposure is not adoption. The source datasets measure how exposed tasks are, or how much usage was observed when they were compiled; adoption barriers, regulatory constraints and organisational inertia mean real-world change typically lags technical capability. Figures also describe the moment each dataset was published, not today.

We maintain a transparent record of major data updates and methodology changes.

The works listed here are among the sources and supporting literature behind the figures on this site. Each figure on an occupation page names the dataset it came from.