The idea
Jobs rarely vanish in one go. Tasks shift, one at a time. So every job here is broken into the tasks it's made of, and each task is marked by who does it today.
What the labels mean
- Human: no regular AI use or robot deployment found for this task.
- With AI: people regularly use AI to help with this task.
- AI: in real usage, AI mostly does this task on its own (at least 75% of the time it shows up).
- Robot: a robot has done this task in a real workplace built for people, or out in the open. Robots that only work in factory cells built around the machine are noted separately.
The 8-hour shift
Each task has an estimated share of a typical worker's day. These shares come from Anthropic's "What work can robots do?" study, where an AI model (Claude) estimated how working time splits across each job's O*NET tasks. They are informed estimates, not measured hours, and real jobs vary a lot. That's why you can untick tasks or set your own minutes with "Adjust my times".
When you untick tasks, their time is spread across the tasks you still do. When you set your own minutes, nothing is rescaled: any time you don't assign shows as unassigned.
How sure is each label?
Every task shows the kind of evidence behind it:
- Observed in AI use: the task shows up regularly in real AI conversations (Anthropic Economic Index, usage from August and November 2025). This shows people use AI for the task. It doesn't show the work was removed from anyone's job.
- Robots deployed: a cited robot does this task in real, ongoing deployments.
- Robots sold commercially or Demonstrated only: the robot exists as a product or has been shown working, but wide use isn't documented.
- Robots only in factory cells: robots can do it only in setups built around the machine, so the task stays Human here.
- No AI or robot evidence found: nothing turned up in the data. That isn't proof the task is safe; it may just be rare in the data.
A task gets one label. Where both apply, AI use takes priority over robots.
Company news stages
Announced, Pilot, Rolling out and Widespread describe what a company actually did. Pulled back marks a reversal. Plan and Forecast (shown with dashed borders) are predictions or internal targets, which may never happen.
Where the data comes from, and how current it is
- Jobs, tasks, job titles, skills and preparation: O*NET 30.1, U.S. Department of Labor/ETA, CC BY 4.0.
- AI use by task: Anthropic Economic Index, "Labor market impacts" (usage data from August and November 2025), CC BY.
- Robots by task, time shares and robot sources: Anthropic Economic Index, "What work can robots do?" (September 2026), CC BY 4.0.
- Employment and median pay: BLS Occupational Employment and Wage Statistics, May 2025.
- Job outlook: BLS Employment Projections, 2025 to 2035.
- Company news: checked by hand from published sources, updated as stories appear.
Limits
All job, pay and outlook data is for the United States. The same job can look quite different in another country, or even another city. AI usage data comes from one AI provider's conversations, so it shows how people actually use AI, not everything AI could do or how every employer works. Robot ratings record what robots have done somewhere, not how common they are. Employment figures for closely related jobs are sometimes shared, because BLS groups them under one code. None of this predicts a date when a job disappears.
Open data
All job data is free to reuse with credit: jobs.json.