How the Job Check works

The Job Check covers 923 US jobs. For each job it shows up to 10 main tasks, most important first, and what four studies say about how AI touches them. The job list, the tasks and 40,396 job titles for the search come from O*NET 31.0, the US Department of Labor’s database of jobs (August 2026).

Each task gets one label, from the strongest evidence down:

• People already use AI for this at work: Anthropic saw enough real work chats with Claude for this task.
• AI could speed this up a lot: in 2023, GPT-4 rated whether an AI model alone could do the task in half the time or less, at the same quality. The study’s authors checked these ratings against human raters.
• AI could speed this up, with extra tools: the same, but only with software built around the model.
• Mostly human for now: no big speed-up expected, and no real use seen.

• Already used at work (Anthropic, 2026-03-05): the share of a job’s working time spent on tasks that Claude is already used for at work. Tasks AI fully does count fully; tasks where AI helps a person count half. A task counts only after enough work use, so a zero often means too little data.
• Could speed up (Eloundou, Manning, Mishkin and Rock): the share of a job’s tasks an AI model could do in half the time, counting tasks that need extra tools as half. These ratings are from 2023, before today’s models.
• Help in Copilot chats (Microsoft Research (Tomlinson et al.), data from 2024): which of the job’s work activities came up in Bing Copilot chats, how often the AI completed them, and how big a part of the work it handled. We show where the job ranks among all jobs.
• Robots (Anthropic, 2026-09-30): for the physical tasks of a job, whether a robot has been shown to do them, in settings from purpose-built factory cells to private homes. We show it only for jobs with physical tasks that robots can do.

• Exposure is not job loss. Anthropic’s researchers found no general rise in unemployment in the jobs where AI is used most. For young people aged 22 to 25, the rate of finding a job in them was about 14% lower than in 2022, a result the authors call only just significant. A job can change a lot and still grow.
• A zero is not “safe”. Some jobs show no AI use because too few of their tasks appeared in the data.
• It is US data. The tasks describe US jobs, and the use data comes mostly from US users. Jobs elsewhere are often similar, but not the same.
• It is not advice. It shows evidence and options, not what you should decide about your career.

• O*NET 31.0 Database, US Department of Labor, Employment and Training Administration (USDOL/ETA), 2026-08. Licence: CC BY 4.0.
• Anthropic Economic Index: labor market impacts (observed exposure), Anthropic, 2026-03-05. Licence: CC BY.
• Working with AI: Measuring the Applicability of Generative AI to Occupations (v1.1), Microsoft Research (Tomlinson et al.), 2025-12. Licence: CC BY 4.0.
• GPTs are GPTs: labor market impact potential of LLMs (Science, 2024), Eloundou, Manning, Mishkin and Rock, 2023. Licence: MIT.
• What work can robots do? (robot exposure index), Anthropic, 2026-09-30. Licence: CC BY 4.0.

This page includes information from the O*NET® 31.0 Database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA), used under the CC BY 4.0 license. We have modified this information; USDOL/ETA has not approved, endorsed or tested these modifications. O*NET® is a trademark of USDOL/ETA. The GPTs-are-GPTs labels are used under the MIT License, © 2024 OpenAI.

Found a mistake? Write to editor@siliconainews.com, and we will correct it. Every correction is listed at the end of the Job Check page.