Method

How the Job Exposure Index works

The index scores 923 US occupations from 0 to 100 on how much of their work AI could speed up. Here is exactly where the numbers come from, and what they do and don't mean.

1. The tasks

Every occupation starts from its task list in O*NET, the US Department of Labor's database of what people in each job actually do. For example, accountants "prepare detailed reports on audit findings". In total the index covers 19,265 tasks.

2. The ratings

Each task was rated in the study "GPTs are GPTs" by Eloundou, Manning, Mishkin and Rock (OpenAI and the University of Pennsylvania, 2023), which used GPT-4 to rate every O*NET task, checked against human raters. We use their published ratings unchanged.

3. Three kinds of task

A task counts as exposed if a large language model could cut the time it takes by at least half, at the same quality.

AI can speed this up on its own. A chat assistant alone could do it at least twice as fast.
AI can speed this up with extra software. It needs software built around the AI, such as access to company data or tools.
Little or no AI speed-up. Physical work, or tasks where AI wouldn't save meaningful time.

The score counts the first kind in full and the second at half weight: score = (direct tasks + ½ × software tasks) ÷ all tasks × 100. This is the study's main measure, called β.

4. Levels

Very high (80–100)28 jobs
High (60–79)69 jobs
Elevated (40–59)323 jobs
Moderate (20–39)221 jobs
Low (0–19)282 jobs

6. Jobs and pay

The number of US workers and median annual pay come from the Bureau of Labor Statistics' Occupational Employment and Wage Statistics, May 2025, the latest release. It refreshes automatically when BLS publishes a new year. Where BLS only reports a broader group, we show that group's figures and say so.

5. Real-world use

The exposure score says what AI could do. To show what people are actually doing with it, each task is checked against the Anthropic Economic Index, which maps millions of anonymised conversations with Claude to the same O*NET tasks. The current data covers April–May 2026 (release 2026-06-26); 3,709 of the 19,265 tasks appear in it. A task counts as "seen in real AI use" when it appears in the published data. It reflects one AI assistant, so tasks done with other tools won't show up, and a task that doesn't appear may still be done with AI elsewhere.

A GitHub job checks for new Economic Index and BLS releases every week and rebuilds the site when they change.

What the score doesn't tell you

Exposure is not job loss. A high score means AI could change how the work is done. Whether that leads to fewer jobs, different jobs or more productive ones depends on employers, demand and regulation.

The exposure ratings reflect GPT-4 in 2023. AI has improved since, and tools now do more on their own. The real-world use data above is refreshed as Anthropic publishes it, and the weekly report covers what changes.

Jobs vary. Two people with the same job title can do very different work. Use the task list on each job page to see which parts apply to you.

Sources and licences

Task data: O*NET database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA), used under the CC BY 4.0 licence. Exposure ratings: Eloundou, T., Manning, S., Mishkin, P. and Rock, D. (2023), "GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models", data published under the MIT licence. Real-world use: Anthropic Economic Index, 2026-06-26 release, used under CC BY. Employment and pay: BLS OEWS, May 2025 (public domain).