Is AI taking jobs - or making them?

If you're mid-career and wondering whether your job survives this, or watching your kids head into all of it, you've probably noticed that everyone has an opinion and nobody has receipts. So this page holds to what can be measured, from two directions at once: what AI is observed doing at work right now, and what employers are asking people for. Neither is a prediction. Both update as the data does.

So, what are we looking at, exactly?

Start with what "exposure" means, because it's simpler than it sounds. The US Labor Department keeps an official list of what every job in America involves, task by task - about 18,000 tasks across 756 occupations. Anthropic, the company behind the Claude AI assistant, took millions of real (anonymised) conversations with Claude and matched each one to the task being worked on. An occupation's exposure is the share of its official tasks where that use shows up. It measures where AI is doing work today - not whether anyone lost a job doing it. The grey row is the average across all 756 occupations (7.7%), and most of the table's story is how far the top sits above it.

Is it taking jobs, then?

What the usage data shows is concentration, and the concentration is extreme. Of 756 occupations, just 11 have AI working across more than half their tasks - and 505 sit under 5%. A computer programmer's work is 75% covered; a registered nurse's, 6%. So the fair reading of the moment is that AI is deep into a recognisable band of work - writing, coding, analysis, records, customer conversations, the jobs done at a screen with language - and barely present in work done with hands, presence, or a licence. Whether that concentration spreads is the question the next releases of this data will answer, and this page updates with each one.

And the making side?

Turn from what AI does to what employers want, and the demand is easy to count: 1 in 18 of the 26,862 live listings on our own index now names AI as a required skill - and 50% of those are not engineering roles. 310 of them are sales jobs, with marketing, finance and HR close behind: ordinary jobs quietly acquiring an AI requirement, the way "computer literacy" once spread through job descriptions. The index also carries titles that didn't exist a few years ago - "AI deployment strategist", "forward deployed engineer" - roles invented to put these systems into companies.

What a jobs board can and can't see

A listings index counts what employers post, so a role that automation absorbed doesn't show up as a headline loss - it just stops being advertised. That blind spot is exactly why we track the index over time rather than reading it as a snapshot: month by month, you can watch AI requirements spread into ordinary listings, watch new titles appear, and watch older ones thin out - and that trend shows the change arriving well before it lands in official statistics. For the occupation-by-occupation long view, the Bureau of Labor Statistics' projections remain the standard reference, and they pair well with the exposure table above: one says where the jobs are heading, the other says where the work already overlaps with AI.

What these numbers can't tell you

The exposure data observes one AI assistant - Claude - so treat it as a floor on AI-at-work, not a census; people use other tools, and plenty of use never maps neatly to a task list. Observed use is not job loss: a task AI touches may take a person plus AI, faster. And our listings index is a live, tech-leaning sample of 843 companies' careers pages - real demand, but nothing like the whole labour market.