AI Exposure (DAIOE)

Construction (overview)

DAIOE combines (i) AI capability progress by subdomain with (ii) occupational work content. First, we assemble yearly measures of AI performance across nine applications (e.g., language, vision, speech), drawing on standardised evaluation benchmarks and leaderboards. Second, we map those AI abilities to the skills and abilities that define each occupation. Finally, for each occupation–year, we take a weighted aggregation (a “match” between capability growth and occupational requirements) to form the exposure index and its nine sub-indices as well as two off-springs (the genAI- and non-genAI versions).

Formal details

Capturing AI capability progress

We track yearly progress in nine AI subdomains—games (abstract strategy, real-time), vision (recognition, comprehension, generation), language (reading comprehension, language modelling, translation), and speech recognition—using 140 benchmarks that have been used to test model performance in AI research (sources: EFF and Papers With Code). For each benchmark, we derive a frontier curve, indicating the state-of-the-art at a given time. Metrics are harmonised so changes are comparable, then summarised each year to capture shifts in the technical frontier.

From capabilities to occupations

Subdomain progress is mapped to 52 O*NET abilities via the Felten et al. (2018) 9×52 matrix. Ability-level exposure is aggregated to occupations with O*NET importance×level weights, yielding an annual exposure change and nine parallel sub-indices. To pay attention to interpersonal roles, we apply a calibrated social-skills discount and then cumulate the annual exposure changes from 2010– to form the DAIOE panel.

Interpretation

DAIOE measures the potential applicability of AI to occupational content, over time—not adoption or automation probabilities. Real-world effects, e.g., depend on the development and adoption of AI applications, complementary investments, organisational choices, and policy.

AI sub-domains

• ♟️ Abstract strategy games — tests long-horizon planning & search.

• 🎮 Real-time video games — adaptive control under time pressure.

• 🖼️🔎 Image recognition — identifying objects & patterns in images.

• 🧩🖼️ Image comprehension — reasoning about what an image depicts.

• 🖌️🖼️ Image generation — creating or editing visual content.

• 📖 Reading comprehension — extracting and using information from text.

• ✍️🤖 Language modelling — generating and transforming text.

• 🌐🔤 Translation — converting meaning across languages.

• 🗣️🎙️ Speech recognition — turning audio into text reliably.

The original paper

The DAIOE is developed, explained, validated, and applied in the most current version of the paper by Engberg et al. (2025).

Suggested citation: Engberg, E., Görg, H., Lodefalk, M., Javed, F., Längkvist, M., Monteiro, N., Kyvik Nordås, H., Pulito, G., Schroeder, S., & Tang, A. (2024). "AI Unboxed and Jobs: A Novel Measure and Firm-Level Evidence from Three Countries" (IZA DP 16717).

DAIOE data

This Google Drive folder contains the index files and documentation (2010 to latest available year).

We provide DAIOE in five different "flavours", each one according to a specific occupational classification: international (ISCO-08), US (O*NET-SOC 2010 and O*NET-SOC 2010), and Sweden (SSYK 96 and SSYK 2012).

Each "flavour"--aka folder--contains the DAOIE in three file formats, for convenience, namely: Excel, .xlsx; Stata, .dta; and general, .csv.

Each file has the following columns/variables:

• Col A: the occupational code

• Col B: the occupational title

• Col C: the year

After this follows two groups of the DAIOE index. The first group is the actual index, in Col's D---N. The index is a score which enables comparisons of AI exposure between occupations and over time; a one-unit change in the index does not have a specific interpretation.

The second group is a version with the percentile rankings of occupations within the year, enabling comparisons of where an occupation is in the distribution of AI exposure, and this is in Col's O---Y.

Both groups of the DAIOE have the same structure. In the index value version (percentile ranking version), we provide the following:

• Col D (O): the general or overall DAIOE index (ranking).

• Col's E---M (P---X): the nine sub-domain DAIOE indices (ranking).

• Col N (Y): the generative AI DAIOE index (ranking).

Note: A few initial rows of each file are empty, corresponding to armed forces occupations or legislators, where data are lacking for computing the DAIOE, because they are not observed in O*NET.