Download the aggregates · read the caveats · use the story angles
Everything Singulariki renders is built from public datasets and published studies. This page
packages the cross-occupation aggregates for download, documents every column, lists every source
and its exact release, states the caveats that must travel with the numbers, and offers
source-cited story angles. Built for journalists and researchers — reuse it, but keep the
honest frame attached.
Our unified AI-exposure index per U.S. occupation — the average of two published studies (OpenAI “GPTs are GPTs” task overlap + Felten/Raj/Seamans AIOE), percentile-ranked across all scored occupations.
Data dictionary & redistribution
Aggregate we computed from public O*NET/BLS data and two published, openly-available studies. Freely reusable with attribution.
Column
Description
title
Occupation title (O*NET-SOC).
soc_code
O*NET-SOC 2019 code.
employment
BLS OEWS May 2024 national employment.
median_wage_usd
BLS OEWS May 2024 national median annual wage.
exposure_value
Unified exposure index (0–100), higher = more task overlap.
exposure_percentile
Rank of exposure_value across all scored occupations.
exposure_band
Low / Moderate / High band of the percentile.
observed_use_percentile
Microsoft “Working with AI” activity-overlap percentile (observed, distinct from potential).
teleworkable
Dingel–Neiman (2020) task-feasibility flag for remote work.
BLS Employment Projections per occupation: projected employment change and annual openings over the 2024–2034 decade. A labor-market projection, independent of AI.
Data dictionary & redistribution
U.S. Bureau of Labor Statistics public data. Freely reusable with attribution.
Column
Description
title
Occupation title.
soc_code
SOC code.
employment_2024_thousands
Base-year employment (thousands).
employment_2034_thousands
Projected employment (thousands).
change_percent
Projected percent change 2024–2034.
annual_openings_thousands
Projected average annual openings (growth + replacement), thousands.
Our band/score aggregate of the ILO 2025 global exposure index, by international ISCO-08 occupation. For the full task-level scores, see the ILO working paper linked in the dictionary.
Data dictionary & redistribution
Aggregate derived from Gmyrek, Berg, Bescond et al. (2025), ILO. Cite the ILO source; raw task-level data is published by the ILO — we link rather than rehost it.
Column
Description
title
ISCO-08 occupation title (international).
isco_code
ISCO-08 unit-group code.
major_group
ISCO-08 major group.
task_count
Number of task statements scored.
mean_exposure_2023
Mean task exposure 0–1 (2023 model snapshot).
mean_exposure_2025
Mean task exposure 0–1 (2025 model snapshot).
mean_delta_2023_2025
Change in mean exposure 2023→2025.
exposed_task_share
Share of tasks in an exposed band (0–1).
Caveats that must travel with the numbers
Exposure is task overlap, not automation. It measures how much of an occupation's tasks today's AI can assist — not jobs lost, not a forecast.
Potential ≠ observed. Exposure (what AI could assist) and observed AI use (what shows up in assistant logs) are different measurements and often disagree.
BLS projections are independent of AI. The 2024–2034 outlook reflects labor-market trends, not an AI-impact forecast.
SOC↔ISCO is approximate. U.S. roles are matched to the international gradient through a many-to-many crosswalk — the closest correspondence, not an identity.
Every figure names its source + release. See the source list below and the methodology page for the full detail.
The jobs a famous 2013 study rated safe from automation — teachers, customer-service reps, managers — are exactly the cognitive, language-heavy roles most exposed to today's AI. The ones it called doomed (drivers, laborers, cooks) are now among the least exposed.
The most AI-exposed quartile of work accounts for roughly a quarter of all projected annual U.S. openings — concrete evidence that high task overlap is not the same as work disappearing.
A model on its own touches about one task in seven; once software is built around it, that jumps roughly four-fold. The disruption story is about tooling and workflow, not raw model capability.
On the ILO's international gradient, clerical, data, and text work is hottest; web and software work moved fastest between 2023 and 2025; physical and care work remains human ground.
AI exposure rises with how textual and screen-based the work is — high in $100k management and $43k customer service alike, low in both $30k food service and $94k bedside nursing. It does not track pay.
Weight each industry's exposure by its real employment mix and finance/insurance sit near the top while construction and landscaping sit near the bottom — the same disruption pressure, distributed by who works where.
