The AI Exposure Atlas

One classification, four countries. Every country publishes occupations in its own scheme — PSCO in Pakistan, NCO-2015 in India, SOC 2020 in the UK, SOC 2018 in the US. All four are derived from ISCO-08, and the ILO republishes each national labour force survey on that common spine. This atlas uses that harmonised series at the 2-digit level, which gives the same 40 occupational groups everywhere and makes the comparison real rather than rhetorical.

Exposure is measured, not guessed. Scores come from ILO Working Paper 140 (Gmyrek, Berg, Kamiński et al., May 2025), which assessed the automation potential of 29,753 occupational tasks using 52,558 human ratings from 1,640 workers, validated against an international expert panel and extended to all 427 ISCO-08 4-digit occupations. A 2-digit group’s score is the mean of the detailed occupations inside it.

The index is versioned, and it moved upward. Between the ILO’s 2023 and 2025 revisions, scores rose across strongly digitised professional and technical work, not just clerical — two further years of capability moved the frontier into jobs the earlier revision treated as safe. This atlas uses the 2025 scores throughout and keeps the 2023 ones beside them, so the shift is not just an assertion: Change since 2023 is a metric you can colour the map by, and the detail panel reports the movement for whichever occupation is selected.

A check on the aggregation. The ILO reports that 34% of employment in high-income countries falls in an exposed occupation. Recomputing that from this pipeline gives 34.3% for the UK and 33.9% for the US — so the aggregation reproduces the authors’ own published result.

Remote work and freelancing. Teleworkability follows Dingel & Neiman (2020), who score occupations on whether their tasks can be performed from home. They publish results for 86 countries; India is not among them. Because their measure is applied at ISCO 2-digit — the same spine used here — their occupational weights can be recovered by non-negative least squares from the 84 published countries for which ILOSTAT also reports a 2-digit distribution. The recovered weights fit those countries with R² = 0.99 and a mean absolute error of 1.8 percentage points, and reproduce Pakistan at 13.1% against a published 13.5%. Applying them to India’s 2025 distribution gives 13.5%. Online-freelance domains are the six tracked by the Online Labour Index, mapped to the ISCO-08 4-digit occupations that constitute that work and scored on the same ILO index.

Pay and informality. Median monthly earnings are ILOSTAT’s 2025 series in local currency, published at ISCO 1-digit — so every tile within a major group shares a value, and the pay layer is coarser than the exposure layer. Informal employment rates are published by occupation for Pakistan and India only; the concept is not measured the same way in UK and US statistics, and the layer is disabled rather than faked for them.