DAIOE: how exposed is each job to AI?

No. DAIOE measures how applicable AI’s advancing capabilities are to what an occupation does. Whether that changes jobs depends on adoption, complements and organisational choices.

No. DAIOE captures potential applicability. If, when and how AI is applied in an occupation depends on the cost of acquiring, adapting and using the technology, the institutional setting (laws and regulations), and more.

DAIOE is dynamic and grounded in measured AI benchmark progress: the score moves year by year as AI capabilities actually advance, rather than scoring each occupation once. It tracks separate capability subdomains over time and down-weights socially intensive work (see below).

The index has no natural units: a value is meaningful only relative to other occupation-years. For cross-sectional standing, use within-year percentiles. One caution: v1.0.0’s rank columns can order occupations with identical values arbitrarily, so where ties could matter compute the tie-safe midrank instead (one line: 100 × average rank within year / count); from v1.1.0 it ships precomputed. Percentiles are ordinal, so they do not show how much exposure has grown; for level and growth, rescale as in the next question.

Yes: divide by the highest value in the frozen 2010 to 2023 panel you use and multiply by 100 (score_rel_max = 100 x value / max). A score then reads as per cent of the most exposed occupation-year, and growth survives: an occupation at 31 in 2015 and 62 in 2023 doubled its exposure. Keep the 2010 to 2023 maximum as the denominator for later vintages too; values above 100 then mean exposure beyond the frozen-window peak. We document this transformation rather than shipping it, so the raw values remain the single citable version; details are in the technical documentation below.

Cognitive, non-physical roles with lower social-interaction intensity, often white-collar, tend to rank higher.

Yes, relative to less exposed firms in the same industry and year. In the Swedish, Danish and Portuguese registers, clerical support employment falls with exposure within firms. In Sweden, the case studied in depth, firms employed 7.1 per cent fewer clerical workers per within-year standard deviation of exposure, weighted by employment. In Denmark, in Portugal and for the typical Swedish firm, the pre-period estimate has the opposite sign. Total employment shows no detectable decline through 2023. See the paper for the full evidence and its limits.

Through a down-weighting built into the measure. The abilities that generate the exposure score capture how applicable AI’s measured capabilities are to a job’s content, but they do not fully capture how interpersonal that content is, and the more a job consists of human interaction, the smaller the part of it AI’s capabilities apply to. DAIOE therefore scores social intensity separately: O*NET rates how important interpersonal skills are in each occupation and at what level the job requires them, and those ratings combine into a score that dials exposure down as it rises. Without such an adjustment, clergy rank as strongly exposed, yet the substance of their work is sustained human relationships, where the measured capabilities apply least. The adjustment tempers exposure rather than erasing it, and the paper shows the results do not hinge on how strongly it is applied.

Capability areas across games, vision and language, each tracked from its own public benchmarks. In the frozen 2010–2023 index: abstract strategy games and real-time video games; image recognition, image comprehension and image generation; reading comprehension, language modelling, translation and speech recognition. They aggregate into the overall index, and language modelling and image generation also combine into the generative-AI sub-index; the paper uses these. The 2025 vintage also publishes second-generation composites in standardised units, over all thirteen applications and over the four generative ones, and this page uses those. Later vintages add subdomains as credible benchmarks emerge, and each vintage documents its own set.

Yes. Combine occupation-level DAIOE with local workforce structures (firms, industries, regions) to assess exposure distributions.

Yes, two. The original generative-AI index combines language modelling and image generation and is the one the paper uses; its membership never changes. From the 2025 vintage, a second-generation generative composite adds conversation and software engineering, in standardised units, and this page uses it for the generative view.

Cite the release and the paper, and name the object you used. The scores carry a DOI per version (v1.1.0: 10.5281/zenodo.22551410; v1.0.0, which holds the frozen index the paper estimates on: 10.5281/zenodo.21873968; all versions: 10.5281/zenodo.21873967). A release contains several objects, so cite the DOI plus the object (frozen 2010–2023 index, 2024 refresh or 2025 vintage) plus the filename you loaded; the repository's CITATION.cff carries the machine-readable form.

A vintage is one released version of the dataset, labelled by its coverage window; vintages are separate, citable objects and are not interchangeable. Use the frozen 2010–2023 index to replicate or compare against published work, and the newest vintage (now the 2025 vintage) for current analysis. The frozen window is carried cell-identically in every later vintage: published values never change.

The bundle ships the frozen 2010–2020 moments (standardisation_moments_v1.csv: mean and standard deviation per classification and column), so standardising is one line, z = (value − mean) / sd, and cannot be done on the wrong window.

The repository README opens with quick starts in Python and Stata: the full ranking of occupations with percentiles, top and bottom ten with titles, and one occupation's exposure path over time. A code–title lookup for the Swedish classifications ships in the bundle.

Yes, by annual vintages, with published values never changing: every release verifies cell by cell that the frozen window is identical, and all basket and membership changes take effect at annual chain points, the convention of chain-linked official statistics. The 2025 vintage was released as v1.1.0 on 8 October 2026, adding agentic task execution and mathematical and scientific reasoning, with companion composites in standardised units.