AI Job Exposure Checker

1. Tasks. Each occupation’s task list comes from the U.S. Department of Labor’s O*NET database: 923 occupations and about 19,000 task statements.

2. Exposure. Every task was rated in “GPTs are GPTs” (Eloundou, Manning, Mishkin & Rock; arXiv 2023, Science 2024) by trained human annotators and, separately, by GPT-4. The question was whether a large language model could cut the time the task takes by at least half at the same quality. The ratings are not exposed (0), exposed with extra LLM-powered software (0.5) or directly exposed (1). Your score averages the human and GPT-4 ratings across your tasks, with core tasks counted double (the paper’s β measure), then scales it to 0–100.

3. Buckets. Automatable now means both raters judged the task exposed, at least one said an LLM alone is enough, and either GPT-4 rated it for significant or complete automation or Anthropic observes it in real AI usage. AI-assisted means AI can speed the task up, but a person stays in the loop. Human-led means neither rater found meaningful exposure (or only one did, and only with extra tools).

4. Real-world use. The “Seen in real AI use” task flags and the occupation-level “Seen in real AI use today” figure (Anthropic’s “observed exposure”) come from the Anthropic Economic Index Labor market impacts data (Massenkoff & McCrory, 2026), which is based on how Claude is actually being used at work.

5. Skills and moves. Suggested jobs are O*NET related occupations that score at least 8 points lower, and usually at the same or a higher Job Zone. Skills are the O*NET skills those jobs need more than yours does, limited to skills whose importance doesn’t rise with AI exposure among jobs at the same Job Zone.

Limits. Exposure measures what AI could speed up, not whether your employer will cut jobs. The ratings reflect early-2023 AI capabilities, so newer models and agents probably push many scores higher. Task lists describe the typical U.S. worker, not your specific role. Treat the result as a conversation starter, not a forecast.

Sources & licenses: O*NET 31.0 Database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA), used under CC BY 4.0. It was modified here: tasks were matched to exposure ratings and scores were computed. USDOL/ETA has not approved, endorsed or tested these modifications. O*NET® is a trademark of USDOL/ETA. · Eloundou, T., Manning, S., Mishkin, P., & Rock, D. “GPTs are GPTs”, Science 384 (2024); data from github.com/openai/GPTs-are-GPTs (MIT License, © 2024 OpenAI). · Anthropic Economic Index, labor_market_impacts (Massenkoff & McCrory, “Labor market impacts of AI”, 2026), data released under CC-BY.