Questions people ask about AI and jobs

Probably not all of it, and the honest answer is job by job. This index scores every task in 923 US occupations and reports three numbers: the share a current AI system can produce with little standing in the way, the share it can produce but cannot finish alone, and the share it cannot produce at all. The median occupation has under a third of its task load in the first group. Look up your own job and you will see which of your tasks are in which group.

The ones held in place by hands, presence or liability. Dishwashers, roofers, meat packers, anaesthesiologists and heavy equipment mechanics all sit in the untouched band, and they have almost nothing in common except that the work happens to physical things in physical places, or somebody licensed has to sign for it. The full list is at /lists/jobs-ai-cannot-do.

Telemarketers, desktop publishers, proofreaders, data entry keyers and title examiners lead the index. What they share is that the whole work product is text or a form, produced from information somebody else supplies, and checking it is cheaper than doing it.

No, and this is the distinction the index exists to protect. Exposure measures what current systems can produce and what structurally stands in the way. It says nothing about whether an employer will adopt the tool, how fast, at what cost, or what the law will allow. Adoption is a separate question with separate evidence, and this index does not model it.

Three things. The unit is a single standardised work activity rather than a whole occupation, so a score can be argued with at the level where the work happens. Every release states the date of the AI capability it was scored against, so a figure can be checked rather than believed. And the index publishes its own inter-rater reliability: eight raters agreed exactly 84.9% of the time on a shared calibration set, and within one point 100% of the time.

Task and activity statements come from O*NET 31.0, the US Department of Labor's occupational database: 923 occupations, 18,838 tasks, 2,087 detailed work activities. Pay and employment come from the Bureau of Labor Statistics Occupational Employment and Wage Statistics, May 2025. The ratings against the rubric are ours, and they are the part open to challenge.

Exposed is work a generally available AI system can produce today with little structural friction. Assisted is work it can produce but cannot finish alone, because somebody has to be physically present, somebody licensed has to be accountable, the context is not visible to a model, or checking the output costs as much as doing the work. Untouched is work it cannot produce at all. The three shares sum to 1 by construction.

Every quarter. Each release carries an identifier and a capability reference date, and the numbers move as systems improve. That is why every figure on the site is published with its release: a number without one cannot be checked against anything.

The published occupation figures are free under CC BY 4.0. Attribute the publisher, link the page you took the figure from, and quote the release identifier. The task level ratings and the quarterly deltas are licensed separately. The full terms are at /terms.

Yes. There is an MCP server at /mcp with five read-only tools, a JSON API at /api, an agent-to-agent endpoint at /a2a, and a markdown version of every occupation at /md. No key and no sign-up. The details are at /agents.

Not yet. The task statements, the pay and the employment counts are all US federal data, so the scores describe US occupations. The rubric itself is not country specific, and the same method would apply anywhere the task data exists.

A.I.T. Multiverse Consulting Ltd, a small IT company in Nicosia, Cyprus. The site is paid for by advertising, by affiliate arrangements with training providers, and by licensing the underlying ratings. None of those parties can change a score: the ratings are published with the rater and the date attached, and the whole method is open to being argued with. How the money works is set out at /disclosure.

Write to [email protected]. If the answer belongs on this page, it goes on this page.