How we worked this out

Used for  Master occupation list

Covers  Global

Updated  Stable

All twenty occupations on this site have had their four-digit unit group verified against the real ISCO-08 hierarchy, as carried in the ILO dataset. Two of them share one: product design and graphic design are both 2166.

Used for  Global AI exposure scores

Covers  427 four-digit ISCO-08 occupations

Updated  Irregular (2023, 2025)

Ingested and in use. 427 four-digit ISCO-08 unit groups and 3,265 scored tasks. The ILO licenses everything published from 3 May 2023 under CC BY 4.0, including datasets, with derivatives and commercial use permitted; the required adaptation disclaimer is shown below. The authors' companion repository carries no licence of its own, which is why the ILO terms are the ones relied on.

Used for  Measured physical and manual work intensity, and how constrained that work is, on every occupation the crosswalk reaches

Covers  Around 900 US occupations and 19,000 task statements

Updated  Quarterly, with the main update in Q3

CC BY 4.0, confirmed against the O*NET distribution terms. 31.0 carries 18,838 task statements and 305,389 work-context rows. Ingested: the mean importance of the four Physical and Manual Work Activities (4.A.3.a.1 to 4.A.3.a.4), rescaled from O*NET 1-5 to 0-100, for 911 O*NET-SOC occupations. Reaches 419 of the 427 ISCO units through the BLS crosswalk chain registered below. Also ingested, for 910 occupations and carried the same way: the perception and manipulation bottleneck of Frey and Osborne (2017, Table I) — Finger Dexterity and Manual Dexterity at their level ratings, Cramped Work Space at its frequency rating, each rescaled 0-100 from O*NET's published scale range and averaged with equal weight — and the level of Assisting and Caring for Others, one of their social-intelligence variables. The variables and scales are theirs; the equal weighting is this site's; their probabilities of computerisation are not used.

Used for  Carrying the Australian augmentation/automation split onto the global classification

Covers  1,345 correspondence rows, reaching 414 of the 427 ISCO-08 unit groups

Updated  As published

Replaces the BLS SOC crosswalk originally registered, which maps ISCO-08 to the 2010 SOC while O*NET 31.0 is on the 2018 SOC — three hops, compounding error, and BLS refuses non-browser requests. This one is a single hop. Most relationships are marked partial by the ABS, and several ANZSCO units can land on one ISCO unit; those are averaged unweighted because the correspondence carries no employment weights. Every value derived through it is labelled derived, never measured, and carries its confidence.

Used for  Carrying O*NET’s measured physical intensity from US SOC onto ISCO-08

Covers  1123 ISCO-to-SOC edges, 840 SOC 2010 codes, 867 SOC 2018 codes; 419 of 427 ISCO units reached

Updated  Republished only when a classification is revised

Three hops, and the site says so wherever the result appears. An earlier note here argued this chain was not worth building and that twenty hand-mapped occupations were more honest than 427 laundered ones. That was right about the risk and wrong about the remedy: the fix is to carry the uncertainty rather than drop the coverage. Each ISCO unit records how many O*NET occupations were pooled to reach it — median 2, worst 35 — and a tier from that. The mean is taken back up the chain so each ISCO-to-SOC edge counts once, rather than letting an SOC code with many O*NET specialisations outvote one with a single profile. 8 units map only onto SOC residual “all other” codes, which O*NET does not profile; they are left empty. The BLS files are served only to requests carrying full browser headers, and the ISCO file is a 2012-vintage OLE2 .xls, read by a purpose-written parser in scripts/ingest/lib/xls.mjs rather than by adding a dependency with open advisories.

Used for  Occupation descriptions, the job titles people use, and 24 languages

Covers  All 427 ISCO-08 unit groups; 1,641 named occupations; 24 languages

Updated  Periodic major versions

Ingested and in use. The only source here that needs no crosswalk: ESCO is built ON ISCO-08, so every unit group is addressable by URI and all 427 resolved with none missing. Reusable for any purpose under Commission Decision 2011/833/EU; EU-owned content on europa.eu is CC BY 4.0. Taken through the public API rather than the bulk download, because that download is gated behind a form wanting an email address and yields no stable URL — so the ingest would not be reproducible by anyone else.

Used for  Australian layer, and the augmentation versus automation model

Covers  357 ANZSCO v1.3 unit groups

Updated  One-off study

Ingested and in use. The only source in this register that separates AUGMENTATION from AUTOMATION, which is the distinction SCO-02 makes central and which WP140 cannot provide. Its own finding: 349 of 357 occupations are augmentation-led. CC BY 4.0 per JSA copyright terms; required attribution is "© Commonwealth of Australia". ANZSCO v1.3, so the twenty profiles carry hand-made codes and the 427 index is untouched.

Used for  Candidate for the robotic-reach input. Assessed and rejected for that use.

Covers  18,797 tasks across 124 countries; occupation panel at ISCO level

Updated  Unknown

Checked and deliberately NOT ingested. It is the only source addressing the physical and robotic channel, and its transparency is better than most — published prompts, validation materials, a careful data-use notice. Three things kept it out. Its labels were produced by a large language model, which is systematic assessment rather than measurement, and putting them in the same visual language as the ILO expert survey would overstate them. Its occupation-level panel is released only as Parquet; the accessible CSVs are ISCO 2-digit country panels, roughly ten four-digit units per value. And its exposure figures exceed 1.0, so they are an index on a scale this site has not established, not a share. The MIT licence covers the authors code, not the third-party data underneath, which the notice says to check before commercial use.

Used for  Optional real-world usage layer

Covers  542 O*NET-SOC occupations with a collaboration split

Updated  Periodic. The method changed between the March and June 2026 releases.

Ingested as OBSERVED CONTEXT ONLY and never as an input to any score. Published by Anthropic about usage of its own product (DAT-09). Three reasons it is kept out of the scoring: it measures Claude users rather than workers; its two figures are shares of one hundred while the JSA figures are independent potentials, so they are not comparable; and for physical occupations it inverts — electricians read 61% automation here against an assessed 16, because nobody wires a house with Claude, they bring it the paperwork. That inversion is the most useful thing in the dataset and the job pages show it as a warning.

Used for  What the public expects of AI and work. Context only, and never an input to any score.

Covers  42,151 people in 36 countries, surveyed 8 February to 13 May 2026, plus two separate US surveys of 5,119 and 3,488 adults

Updated  Periodic. Comparable questions were asked in 2024.

Quoted, never ingested. It is the only source in this register that measures what people believe rather than what an occupation contains, and it is kept out of the scoring for exactly that reason: a survey of expectations is evidence about expectations. It is registered because the gap it opens is worth naming. In 34 of the 37 countries Pew covers, people lean towards AI meaning fewer jobs; the one source here that separates augmentation from automation finds 349 of 357 Australian occupations augmentation-led. Both findings can hold at once, and neither settles the other, because they are answers to different questions. Pew permits excerpts with a citation and does not permit a report or a dataset being republished in full or substantially in full, so figures are quoted on the page that uses them, the survey microdata is not redistributed here, and the required citation is: "Globally, More People Expect AI to Cause Job Loss Than Growth." Pew Research Center, Washington, D.C. (17 September 2026).

Used for  Where US workers may need to move by 2035, and how hard the move is, as one consultancy models it. Context only, and never an input to any score.

Covers  US occupations from 2025 to 2035, modelled on US Bureau of Labor Statistics and Lightcast data

Updated  One-off report, in a long-running MGI series on automation and work

Quoted, never ingested. It is a model of one country, built by a consultancy that sells advice on exactly this kind of change, and its headline figure is a base case inside a wide range: about 11 million people changing occupation, somewhere between 6 and 16 million. It is registered because it answers a question this site's scores can't. The scores ask how much of a job today's AI could do; this asks where people go when demand for their job falls, and how hard the road there is. Its occupations are US ones. Where a post sets one beside an ISCO unit on this site, the match is made by hand, and the post says so.

Used for  How many US job cuts employers announce each month, and the reasons they give. Context only, and never an input to any score.

Covers  Job cuts announced by US-based employers

Updated  Monthly

Quoted, never ingested. It counts announcements, not people out of work, and it records the reason the employer gives. So when AI tops its list of reasons, that measures what companies say about their cuts. Some of those cuts would have happened anyway, and some cuts that AI did cause will have been put down to something else. It is widely quoted, which is why its numbers turn up in headlines, and why it is worth reading the whole table rather than one line of it.

Used for  How much first-person footage of factory work has been published for training robots. Context only, and never an input to any score.

Covers  Head-mounted video of factory workers: 10,000 hours from 2,138 workers, and about 100,000 hours from 14,228

Updated  One-off releases, each larger than the last

Quoted, never ingested. Only the sizes on the dataset cards are used: hours of footage and how many workers wore the cameras. No footage was downloaded, and none is shown. It is registered because it gives the scale of what is being collected, which is the question a post about filming workers' hands has to answer. Build AI's business is robot training data, so it has an interest in this footage being seen as valuable. The cards are gated behind a contact form; the figures quoted are the ones published on them and repeated in coverage of their release.

Used for  The official US forecast of which occupations grow and shrink. Quoted for comparison, never an input to any score.

Covers  Every US occupation; the 18 AI- and IT-related occupations named in the BLS summary are the ones used here

Updated  Annual, each covering the next ten years

Quoted, never ingested. A projection is a forecast of employment: it weighs demand, cost, trade, demography and what employers say they will do, over ten years. A score on this site is none of those things — it is what the ILO’s raters said about the tasks in a job today. The two are set side by side in one article precisely because they disagree, and the disagreement is the point.

Used for  How many advertised vacancies Australian employers fill, by occupation group. Quoted for comparison, never an input to any score.

Covers  The eight ANZSCO major groups, nationally; the national fill rate was 68.2% in the March quarter 2026

Updated  Quarterly

Quoted, never ingested. The fill rate is the share of advertised vacancies that were filled. It measures how hard a job is to hire for this quarter, which depends on pay, training pipelines, location and much else; it says nothing about AI. In one article it is set beside Jobs and Skills Australia's own automation scores (jsa-genai), averaged unweighted across the unit groups in each major group, because the two line up across the eight groups (a correlation of 0.78) — a pattern, not a cause. Required attribution: © Commonwealth of Australia.

Used for  Monthly US employment by age group and AI exposure, from payroll records. Quoted and redrawn for comparison, never an input to any score.

Covers  Millions of US workers in ADP's payroll sample, by six age groups and five quintiles of occupational AI exposure (GPT-4 ratings), indexed to November 2022

Updated  Monthly

Quoted, never ingested. One article redraws the published employment index for workers aged 22 to 25 and 35 to 40 in the least and most exposed quintiles, and the paper's Table 1 (panels A, C and F), which shows how much of the young-worker gap survives controls for education, interest-rate exposure and remote work. The authors call these early, descriptive indicators, not causal estimates, and note that the gap is sharper in ADP's sample than in national surveys.

Used for  First jobs and pay of US graduates by degree and its AI exposure, before and after ChatGPT. Quoted for comparison, never an input to any score.

Covers  Bachelor's graduates from institutions in the Post-Secondary Employment Outcomes data, about 29% of US bachelor's degrees awarded 2016 to 2024, linked to LEHD earnings records

Updated  One-off working paper

Quoted, never ingested. One article uses the headline estimates (initial employment down five points and initial earnings down 13% for the most exposed tenth of majors), Appendix Table A3 (each degree's 2024 coefficients against 2022, centred on the average degree) and Figure 5. The Figure 5 values are read back from the chart: each sector's change in share is its position, which is exact, and its 2022 share is the area of its circle, which is close. A working paper, not yet through the Bureau's full review.

Used for  First-year employment, earnings and graduate-school enrolment of Texas graduates by their degree's exposure to automation. Quoted for comparison, never an input to any score.

Covers  Graduates of Texas public four-year universities, May 2021 to May 2024 cohorts, from Texas Higher Education Coordinating Board and Texas Workforce Commission records

Updated  One-off article

Quoted, never ingested. Used for three findings: more exposed majors were more likely to go on to graduate school, mostly in the same field (two thirds of returning computer science graduates chose computer science); new master's holders in exposed fields saw similar relative pay declines; and enrolment in more exposed majors fell behind between autumn 2024 and autumn 2025.

Used for  Unemployment of recent US graduates by AI exposure, summer 2026. Quoted for comparison, never an input to any score.

Covers  US bachelor's graduates aged 22 to 25 no longer in school, summers 2022 to 2026

Updated  One-off working paper

Quoted, never ingested. Used for summer 2026 unemployment of recent graduates (7.3%, against 7.8% in 2024), the wider 'sidelined' rate (10.4%, the highest of five summers but not significantly so), and the finding that the gap between more and less AI-exposed occupations was small and not statistically significant.