How we score AI exposure - and what we don't claim
Every score on this site answers one question: what share of this occupation's core tasks can current AI systems demonstrably do today? That's it. It is a capability-and-evidence measure, not a prophecy about layoffs, salaries or the future of any company.
Step 1 · Tasks, not job titles
We map occupations to their official task statements (task lists published by O*NET, the U.S. labor department's occupational database, © O*NET OnLine, CC BY 4.0). Tasks are the honest unit of automation: "drafting reports" or "invoice processing" is something a model either does or doesn't do - a job title is 5–50 such tasks bundled together. 14444 tasks and 14413 task–occupation links are mapped today.
Step 2 · Evidence, not vibes
A continuously running pipeline ingests published evidence: capability research, model releases and updates from the labs themselves (their own RSS feeds), benchmark/agent deployments, hiring signals, and labor-market studies. Each signal is fingerprinted, deduplicated and attributed; 149 evidence itemscurrently back the tables. 20 sources are polled on their own schedules - source list here.
Step 3 · The five score components
- Task automation - proven task-level coverage by AI (the largest component).
- Capability - how strong today's frontier models are on the task category.
- Adoption - real deployments and tool availability, not press releases.
- Labor-market - hiring, wage and postings evidence where it exists.
- Evidence quality - a score without sources is capped, never league-table padding.
Exact weights, versioning and formulas: the fully technical methodology page.
Step 4 · Quality gates
An occupation only appears with a score when: it has an official SOC code, at least 10 core tasks, 100% task-coverage (every core task scored), and at least the site-wide minimum of attributed evidence items. Occupations that fail a gate are visible but unscored. Nothing on the rankings below the evidence floor is ranked - which is why our league tables occasionally exclude jobs you might expect to see.
What we deliberately do NOT claim
- Not a layoff forecast. A 70% score means 70% of listed tasks are demonstrable with AI. Jobs also contain in-person constraints, supervision, liability, trust and preference - none of which are in the score.
- Not a wage or salary model. Median wages are displayed (BLS data) for context; the score does not embed pay.
- Not a single-vendor verdict. Scores cite evidence across providers; page names the models linked to each task, not marketing claims.
- Not silently rewritten. Score changes carry a version and snapshot history(39704 snapshots today); if a score improved because evidence landed, you see the delta.
Use the data
All tables are open: downloads, the public API, per-job embed widgets, and editor@aikillswhat.com for corrections. Corrections are welcome and credited.