A guide to the numbers · 2026 edition
How to read this
The atlas compares work with AI capabilities assessed in 2023. It helps you explore task overlap. It does not predict which jobs will disappear.
A job is made up of many tasks.
Researchers rated whether an AI language model could help complete each task in half the time, at the same quality. Some tasks required additional software built around the model. The score combines those ratings; it is not a percentage chance of losing a job.
The atlas groups occupations into four tiers. These describe the 2023 ratings, not a test of today's tools.
- high overlap
- most tasks overlap in the 2023 AI ratings.
- substantial overlap
- a large minority of tasks overlap.
- some overlap
- some tasks overlap.
- little overlap
- few tasks overlap.
High and substantial together form the headline figure. The dividing lines are choices: a small score difference near a boundary does not imply a sudden change in real-world risk.
Picture 100 Colorado jobs.
About 27 are in occupations in the high or substantial overlap tiers. Each square represents one job; the picture rounds 27.4% to 27.
High + substantial overlap Some + little overlap
What the picture counts
Jobs in occupations with higher task-overlap ratings. In the full estimate, that is 784,367 of 2,860,276 covered jobs.
What it does not count
Jobs expected to disappear, or tasks already automated. The other 73 jobs are not necessarily safe or unaffected. A task rating cannot predict an individual worker's future.
“Covered” means jobs counted in unemployment-insurance records. Most self-employed work is excluded.
Understand the ratings and their limits →County figures are estimates of local work.
Federal data tells us how many jobs a county has in each industry. The atlas uses national staffing patterns to estimate the kinds of occupations inside those industries. It does not directly count each occupation in each county.
Some industry counts are withheld to protect confidentiality. Those gaps must also be filled in. “Employment estimated” on a county page describes this second step. A larger percentage means less local information was available; read broad patterns rather than small differences.
Even “No suppressed industry cells” does not make the county's occupation mix an observed count.
Occupation pages use published statewide employment and wages instead of county estimates. When an occupation has no direct task rating, its page identifies the group-based estimate.
A typical county is different from a typical job.
The maps use Colorado's overall 27.4% as their reference. Larger job markets count more toward this statewide figure. The median county is 21.2%: half of counties lie on either side, regardless of size. “Typical county” bands in the detailed table use that second comparison and are shown without map colors.
Three dates, three different meanings.
- 2023: capability ratings
- Human task ratings from Eloundou and colleagues. Updating employment does not update these ratings.
- 2025: employment
- BLS employment data describes the mix of jobs to which the ratings are applied.
- 2026: atlas edition
- The publication edition brings those sources together. It is not the date of an AI capability test.
You can inspect the original sources, read the full methodology, or download the figures.