What "task exposure" actually measures (and what it does not)
Every serious estimate of AI job risk starts from task exposure: the share of an occupation's individual tasks that a large language model can meaningfully do. It comes from the 2023 research paper "GPTs are GPTs" (Eloundou et al.), which scored thousands of O*NET tasks for LLM exposure using both GPT-4 and human annotators.
Jobs are bundles of tasks
No occupation is one thing. A paralegal does document review (highly exposed), client coordination (barely exposed), and court-filing logistics (somewhere in between). Task exposure measures the bundle, which is why two jobs with the same title can carry very different risk depending on their actual task mix.
Exposure is not the same as replacement
A task being exposed means AI can do it — not that it will, immediately, everywhere. Deployment lags capability because of cost, regulation, trust, and integration. That is why our index blends exposure with the Anthropic Economic Index's real-usage data: capability tells you what is possible, usage tells you what is actually happening.
Related occupations
Accountant (53%) · Customer Service Representative (62%) · Market Research Analyst (56%)