What does “exposure” mean?

O*NET, the federal occupational database, describes each occupation as a list of task statements — 30 for this one, 13 of them “core” tasks and 17 “supplemental.” For the exposure measure (Eloundou et al.), human annotators read each task in 2023 and rated whether access to a large language model could reduce the time a person needed to complete it by at least half, at the same quality:

E1

Access to the model alone was rated as sufficient to meet that time-saving threshold. This does not require the model to work without a person.

E2

The same potential time saving, but requiring additional software built around the model. E2 receives half credit under the study's middle measure. That convention is not a measured probability of adoption or a forecast of hours saved.

E0

Neither route was rated as meeting that threshold. This is not a claim that AI could never assist with the task.

These are 2023 judgments, not a test of today's tools. The atlas uses this published baseline consistently across occupations. The study's exposure rubric explains the ratings and their limits.

Task statements are O*NET verbatim. Tasks marked core are the ones O*NET treats as central to the job; they count double in the score.

Some labels may seem surprising. Operating accounting software is rated E0, while other record-keeping tasks are rated E1 or E2. These judgments concern the additional time saving from language models under the 2023 rubric; they do not measure everything that accounting software can already do. A task list also cannot describe every workplace.

With core tasks counted double, the 30 statements carry 43 units of weight. The score is the E1 share of that weight, plus half the E2 share, scaled to 100:

score  =  (E1 weight + ½ × E2 weight) / total weight
       =  (6 + 7.5) / 43
       =  13.5 / 43
       →  31.4 on the 0–100 scale

This reproduces the atlas's published score at its displayed precision: the fraction 13.5/43, multiplied by 100, rounds to 31.4. The 31.4 on this page and on the occupation's own page is this arithmetic and nothing else. A plain unweighted mean over the 30 tasks would give 31.7; the small difference is the core weighting.

The half weight on E2 is inherited from the source. Counted at full weight, the E2 tasks would put this occupation at 48.8; dropped entirely, 14.0. Every score in the atlas carries the same convention — the methodology flags it wherever tier boundaries are discussed.

Here, 4 tasks receive full credit, 11 receive half credit, and 15 receive none; core tasks count double. That produces 31.4 on a 0–100 index. It is neither a simple percentage of task statements nor the percentage of a worker's week that AI could save. The weights do not measure time spent on each task. The score also cannot tell an employer how many clerks to hire.

A useful next study would compare this baseline with later evidence of tool adoption, changing tasks, hiring, employment, and wages. It should specify the comparison groups, dates, and outcomes before testing them, and keep the 2023 ratings fixed for that comparison. A correlation would not establish that AI caused the change. Similar employment growth would not, by itself, show that the task ratings were wrong: firms could adopt tools while expanding output or changing duties. This atlas currently provides the baseline, not the results of that study.