How we score AI exposure

1. Take each occupation’s 41 Generalized Work Activities from O*NET, with their Importance and Level ratings.
2. Weight each activity by how exposed that kind of work is to today’s AI (the coefficients below).
3. Dampen the score for jobs that need a body in a place — physical presence is structural protection.
4. Rank every career and split them into five bands by fixed quota, so the score always carries information.

• O*NET Generalized Work Activities (Importance + Level) and Work Context — the U.S. Department of Labor’s occupational database.
• The approach follows the published literature on AI task exposure — Felten, Raj & Seamans (AI Occupational Exposure); Brynjolfsson, Mitchell & Rock (Suitability for Machine Learning); Eloundou et al. (“GPTs are GPTs”). We are re-implementing a known method, not inventing one.

For each occupation we compute a raw task-exposure score as an importance- and level-weighted average of its work activities’ coefficients:

raw = Σ( coefficient × Importance × Level ) ÷ Σ( Importance × Level )

An activity like Documenting/Recording Information carries a high coefficient; Assisting and Caring for Others carries a low one. A job made mostly of the former scores high; one made mostly of the latter scores low.

A job that needs a body in a place is structurally protected regardless of its information content — this is what separates a barista from an analyst. From O*NET Work Context we build a physical-presence index (standing, handling objects, face-to-face work, hazards, protective gear) and pull the raw score down by up to 55% for the most physical roles:

exposure = raw × ( 1 − 0.55 × physical-presence )

Context signals used: Spend Time Standing, Spend Time Walking and Running, Spend Time Using Your Hands to Handle, Control, or Feel Objects, Tools, or Controls, Spend Time Making Repetitive Motions, Spend Time Bending or Twisting the Body, Spend Time Climbing Ladders, Scaffolds, or Poles, and more.

The old score put every career in “Moderate” — a field whose value never changes tells you nothing. So we percentile-rank all careers on the damped score and split them by hard quota. No band can hold more than 40% of careers; a build that violates this fails automatically.

This is the one place a human opinion enters the score, so it’s public and arguable. Each of O*NET’s 41 work activities gets a coefficient from 0 (untouched by AI) to 1 (fully exposed), with a one-line rationale. Disagree with one? That’s the point — they’re meant to be debated.

• Exposure is not job loss. A high band means a role’s tasks are automatable — not that the job disappears. History suggests exposed tasks are as often augmented as replaced.
• This is a model, not a prediction. It measures task structure today, using ratings that update on O*NET’s schedule. It cannot see your specific employer, your seniority, or how a given company adopts AI.
• High exposure can be good. If AI does the automatable parts, the human can do more of the valuable parts. Exposure tells you where change is coming, not whether it helps or hurts you.
• Coefficients are debatable. They are our best reading of the literature, not settled fact. They are all listed above precisely so you can push back.