04Work About 3 minutes
Is your job exposed to AI?
Pick an occupation and see three separate measures side by side: how exposed its tasks are to large language models in principle, how much of its work shows up in observed Claude activity, and how many people do the job and what it pays. Neither exposure measure predicts job loss.
- Pick a job
- Make a guess
- Compare three measures
- Explore the tasks
1 Choose an occupation
Loading occupation data.
2 Before you look Your guess
What potential exposure score would you give this job?
The score is the share of its tasks where GPT-4 judged that a language model could cut the time needed by at least half (tasks that also need extra software count half).
3 Three measures, kept separate
Research rating Potential task exposure
Human annotators in the same study gave . Rank .
Observed Claude activity
Rank . Covers one AI system, Claude, not all AI use.
Official statistics Employment and pay
Data: task ratings published 2023; Claude usage samples from August and November 2025 (released March 5, 2026); employment and wages May 2024. How these are measured
4 Example tasks behind each measure
Rated as potentially exposed
Most observed Claude activity
5 Potential exposure versus observed Claude activity
Each dot is one of 756 occupations. The two scores are built in different ways (a simple average of task ratings, and a time weighted measure that gives automated use more weight), so they are not on one scale: read a position as a ranking on each measure, not as a gap to close. Many occupations (411) have an observed score of zero and sit on the bottom axis. Hover over or tap a dot to explore, and click one to select it.
How the spread of AI through everyday work shifts who captures the gains is a central argument of Intelligence Capitalism by Bharat Rao.
Sources and method
- Potential task exposure. Tyna Eloundou, Sam Manning, Pamela Mishkin and Daniel Rock, "GPTs are GPTs: Labor market impact potential of LLMs," Science 384 (2024); working paper 2023. Files data/occ_level.csv and data/full_labelset.tsv at github.com/openai/GPTs-are-GPTs (MIT License). The score shown is the GPT-4 rated beta measure: the share of an occupation's O*NET tasks where a language model alone could cut the time needed by at least half, with tasks that also need additional software counted at half weight. It measures potential exposure, not the likelihood that a job is lost or changed.
- Observed Claude activity. Maxim Massenkoff and Peter McCrory, "Labor market impacts of AI: A new measure and early evidence," Anthropic, March 5, 2026, anthropic.com/research/labor-market-impacts. Files labor_market_impacts/job_exposure.csv and task_penetration.csv in the Anthropic Economic Index dataset, huggingface.co/datasets/Anthropic/EconomicIndex (data CC BY). The observed exposure score counts a task as covered when it is theoretically feasible and appears in enough work-related Claude conversations (Economic Index samples from August and November 2025), gives automated use full weight and augmentative use half weight, and weights tasks by the time workers spend on them. It reflects one AI system only and is not a measure of adoption across the population.
- Employment and wages. U.S. Bureau of Labor Statistics, Occupational Employment and Wage Statistics, May 2024 national estimates, bls.gov/oes (public domain). BLS publishes annual wages at or above $239,200 only as a range.
Method. The three sources are joined on six-digit 2018 SOC codes. All 756 occupations in the Anthropic file match both other sources; together they cover about 91 percent of May 2024 employment, and 75 BLS detailed occupations have no exposure scores. Nothing is imputed. Where O*NET lists several detailed codes for one SOC occupation, we use the base code (ending in .00) when present and otherwise the simple average (24 occupations). The two exposure scores come from different releases and methods and are shown side by side, never subtracted: 411 occupations have an observed score of zero, and 7 show an observed score above the potential score because of the different weighting. Example tasks are illustrative: the first list shows up to three tasks GPT-4 rated as directly exposed (core tasks first); the second lists the tasks with the highest observed task penetration in the Anthropic file, matched by exact O*NET task text (every task matched). Ranks run from 1 (highest) to 756.