Measuring the AI transition. Everywhere.

GAIA maps observed AI adoption, theoretical exposure, and real task performance across 178 countries and 923 occupations.

DOI: 10.5281/zenodo.20320112
Scroll to explore ↓
The coverage
0
countries tracked via Claude.ai behavioral data
The granularity
0
occupations rated on O*NET-SOC standard codes
The surprise
Exposure ≠ performance
GAIA-E exposure →task success ↑
β = 0.00
Higher theoretical AI exposure does not predict higher real task-success. Illustrative scatter; coefficient from the Exposure-Without-Performance analysis.
The divide
Coursework use, by income
31%
Low-income
10%
High-income

The 10 most AI-exposed occupations

Published E1+E2 β scores — the share of tasks where AI could cut completion time by 50%+.

Data vintage: Feb 2026 AEI release

Mean theoretical exposure (E1+E2) — Eloundou et al. 2024.

Source: Eloundou, Manning, Mishkin & Rock (2024), Science 384(6702). Displayed values are the original published β scores.


Work is the leading use case — ahead of personal and coursework

Nearly half of all Claude.ai requests in February 2026 were work-related. Personal use follows at 42%, and educational use accounts for 12% — signaling that AI has already crossed from experimentation into daily professional workflows.

  • 45.2%
    Work
  • 42.3%
    Personal
  • 12.4%
    Coursework

Work is now the single largest way people use Claude.ai — a sign AI has moved past experimentation into everyday professional work, not just personal or school use.


Theoretical AI Exposure (E1+E2)

Share of tasks where AI could reduce completion time by 50%+, as measured by Eloundou et al. (Science, 2024).

Highest E1+E2 exposure
Loading…
Lowest E1+E2 exposure

Source: Eloundou, Manning, Mishkin & Rock (2024), Science 384(6702). Displayed values are the original published β scores.

Browse all 923 occupations →

How people collaborate with AI at work

Claude.ai logs show six distinct collaboration modes. Directive use — giving direct instructions to complete tasks — dominates, followed by task iteration and learning.

32.6%
Directive
User issues explicit instructions for AI to execute
25.6%
Task Iteration
Refining outputs through back-and-forth exchanges
22.4%
Learning
Using AI to build understanding or gain knowledge
11.5%
Feedback Loop
Giving AI structured feedback to improve responses
4.9%
Validation
AI checks or confirms human-generated work
3.0%
None / Other
Unclassified or minimal interaction patterns