← Back to the map

How the AI exposure is measured

The colour of every tile on this map — the AI exposure score, 0–10 — is derived from published exposure research, occupational crosswalks and, where official mappings are unavailable, clearly identified model-assisted mappings. It combines two open, generative-AI-era research datasets, mapped onto each country’s official occupation classification.

Data sources

The exposure index draws on two published research datasets, both openly licensed for reuse:

1. ILO Working Paper 140 — Generative AI and Jobs: A Refined Global Index of Occupational Exposure (2025)

Published by the International Labour Organization (a UN agency). Its Annex Table A1 gives a generative-AI exposure mean (0–1) for the 112 ISCO-08 occupations with meaningful exposure. We use these as the authoritative anchor for the high-exposure band.

Gmyrek, P. et al. (2025). ILO. Licensed CC BY 4.0. · Publication
2. Eloundou et al. — “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of LLMs” (2023)

From OpenAI. Provides a task-based LLM-exposure score (“beta”, 0–1) for ~800 O*NET-SOC occupations, continuous across the whole range. We use it to resolve the low-to-mid exposure band continuously.

Eloundou, T., Manning, S., Mishkin, P., Rock, D. (2023). MIT-licensed. · Dataset

The two 0–1 scales agree closely where they overlap (Data Entry Clerks: ILO 0.70 / Eloundou 0.696; Accountants 0.51 / 0.54), so their raw scores can be combined on the same scale without rescaling.

How each occupation gets a score

  1. Raw 0–1 exposure — ILO takes precedence, Eloundou fills the rest:
    • United States: Eloundou beta by SOC-6 code (group mean where a code is missing).
    • Other countries: national code → ISCO-08 unit group, then the ILO mean if that ISCO group is one of the 112, otherwise the Eloundou beta via the ESCO/O*NET ISCO→SOC bridge.
  2. Global percentile — the raw score is ranked against a single fixed global reference distribution and expressed as a 0–100 percentile. One global anchor for every country, so the numbers stay comparable across borders.
  3. Map colourexposure = round(percentile / 10), a 0–10 scale (green = low, red = high).

Country coverage

Each country’s occupations are mapped to ISCO-08 (or, for the US, to SOC):

MethodCountriesHow
DirectUS · IE · IT · NL Occupation code is already SOC (US) or ISCO-08 (IE/IT/NL) — joined directly.
CrosswalkAU · NZ · DE · UK · CA · ES Official classification correspondence (ANZSCO, KldB, SOC, NOC, and Spain’s INE CNO-11↔ISCO-08) → ISCO-08.
AI-mappedFR · JP · KR No clean official ROME / JSCO / KECO → ISCO table was obtainable, so each occupation is placed onto the official ISCO-08 structure by an LLM (a clearly identified model-assisted mapping), then scored as above. Tagged separately (_llmmap); will be upgraded when the official table is wired in.

About the underlying job data

Tile area is the size of the workforce in each occupation, from each country’s official statistics (see the “Data sources” note in the sidebar of every country page — e.g. Jobs and Skills Australia & ABS, US BLS & O*NET, Statistics Canada, ONS, Destatis, and so on). Salary and workforce figures blend official data with estimates; treat everything as indicative, not advice.

Download the data

The full scored dataset — every occupation in all 17 countries, with its AI-exposure score (0–10), global percentile, workforce size and average pay — is available as a single CSV:

dataset.csv — one row per occupation across every country (country, occupation, official code, category, AI-exposure 0–10, percentile, average annual pay, workforce). Free to reuse with attribution to aijobriskmap.com.

Per-country high-resolution PNG maps are linked from each country page and from the home page.

Read the source papers

Exposure is recomputed from the sources above. Contains public sector information licensed under CC BY 4.0 (ILO) and MIT (OpenAI); adapted by percentile normalisation and crosswalking. This site is independent and not affiliated with, or endorsed by, the ILO or OpenAI.