Not whether your job survives. Which of the tasks in it a machine can already finish in half the time, which would need software building first, and which it cannot touch. 19,265 tasks across 923 occupations, every figure sourced.
Free and public, with no sign-up. The underlying data refreshes and so does this, with the version stated on every page.
Every occupation is analysed and searchable, and 771 of them have a full page of their own. 152 occupations have fewer than fifteen rated tasks and do not get one, because a page with nine rows on it is not worth landing on.
Search 7,953 job titles. You will get the three bands for your occupation straight away, whether or not its full page has been published yet.
This is the opposite of every previous wave of automation. Machinery came for physical work and left thinking alone. This does the reverse.
In work needing the least preparation, 76 per cent of tasks are beyond an AI's reach. In work needing considerable preparation it is 30 per cent. Tasks a chat window can already halve roughly triple across the same range.
The most exposed occupations in the economy are survey research, writing, translation and public relations. The least exposed, with no exposed tasks at all, are pile driving, floor sanding, rebar fixing and dredge operating.
One row per occupational group, plus the People profession, which is not a group in the taxonomy but is who this is written for. Spread is the gap between the most and least protected occupation inside that row, protected meaning the share of its tasks that cannot be sped up at all. It matters: a group with a hundred point spread has an average that tells you almost nothing.
Chapters are being written and published one at a time. Rows without a link have their data here and no chapter yet.
Spread is the distance in percentage points between the most and the least protected occupation in a group, where protected means the share of a job's tasks that cannot be sped up at all. It says whether the row above it can be trusted: a narrow spread means the group average describes the jobs inside it, a wide one means it describes none of them.
Computing and maths has the narrowest spread in the economy at 24 points, which is its own finding: almost nobody in it is sheltered. Business and finance and Arts, design and media both run the full range from 0 to 100.
Every figure measures one thing: whether published research judged that an AI could finish a task in half the time, at the same quality. A task that takes half as long still needs somebody doing it, and no employer was asked what they intend to do about any of it.
It must never be used to choose who to make redundant. An exposure figure is not an objective, fair and consistently applied selection criterion.
Open datasets, joined. Nothing here is scored by a model: exposure comes from published ratings made by human annotators, and the build only counts and joins them. That is the whole reason the numbers can be defended.
Those ratings are then checked against two independent studies built from recorded AI use. Between them they cover 15.2 million real interactions with three different AI systems, across more than 150 countries. None of it is mixed into the figures. It is there to test whether the ratings describe something real.
Exposure ratings from Eloundou, Manning, Mishkin and Rock,
GPTs are GPTs: Labor market impact potential of LLMs, Science 384,
1306–1308, 2024, used under the MIT licence.
This report includes information from the O*NET Database by the U.S.
Department of Labor, Employment and Training Administration (USDOL/ETA).
Used under the CC BY 4.0 license. O*NET® is a trademark of USDOL/ETA.
People Team AI has modified all or some of this information. USDOL/ETA has
not approved, endorsed, or tested these modifications.