Job titles lag behind the work

Try typing “AI red teamer” into the job search on this site. Nothing. It’s a real job, people are hired to do it, and it has its own write-up on the new jobs page, but it has no score. That’s not an oversight. It’s how the plumbing works, and it’s worth understanding.

Everything on this site hangs off an occupation list. ISCO-08 is the master one: 427 unit groups, a stable classification, which is exactly what you want underneath a dataset. Stability is what lets a figure from one source be carried onto a figure from another and still mean something.

Stability has a cost, and the cost is time. A classification that changed every year would be useless for comparison, so it doesn’t — which means it cannot name a job that appeared last quarter. ESCO carries the titles people actually put in job adverts, across more than two dozen languages, and it lags too, because a title has to be common before a classification can responsibly add it.

So there is a window, and in a fast period the window is roughly a few years wide. During it, real work is being done, advertised and paid for under titles that appear in no classification, and therefore in no score on this page.

What that looks like on this site

It looks like a gap between two kinds of page.

Every job page here answers: of the tasks in this occupation, how many are exposed to generative AI? The occupation is the unit of analysis. That is a good question for work that has a name. It is the wrong shape entirely for work that doesn’t yet.

So the new jobs page is built the other way round. It starts from roles rather than classifications — eighteen of them as it stands, and that count is one data file away from being different — and each one is placed by how real its title is, not by how exposed its tasks are:

• Hiring now — people are being hired under this title in numbers a survey or job-market dataset has picked up.
• Forming — named and defined by a serious source, with early job adverts, but titles and duties still vary.
• Early — the work clearly exists; the job title mostly does not yet.

Those three stages are doing the work that an exposure score cannot do here. A score assumes a stable task list to average over. These roles are precisely the ones whose task lists are still being argued about by the people doing them.

Most of it is recombination

The tempting story about new work is that it is new. Read down the list and that is mostly not what is happening. What is happening is recombination: an existing craft plus a new object to point it at.

The data centre technician is an electrician or a plumber whose building happens to be full of racks and cooling loops. The legal engineer is a paralegal who also builds the system rather than only operating it. The expert AI trainer is a nurse, a teacher, an accountant, being paid for the judgement rather than the throughput. The robot teleoperator is a machine operator whose machine is somewhere else.

Each of those pairs an old skill with a new surface, and each one is listed on the new jobs page with the jobs it starts from — links into this site, which is also how the build catches a mistake: name a start-from job that doesn’t exist and the page refuses to build rather than shipping a reader a dead link.

The practical reading of that pattern: the thing that transfers is rarely the tool. It is the judgement about when the tool is wrong, which is why the roles hiring hardest are so often filled by people from the occupation the technology was supposed to replace.

How to read a score in the window

Three habits, which is all the honest advice there is.

1. A score describes an occupation, not your job. Yours is a bundle of tasks, and the bundle is rarely the standard one. The further your actual week is from the classification’s task list, the less the number is about you.
2. A title that does not exist yet has no score, and that is not the same as a low one. Nothing on this site says anything about work it cannot name. Absence is absence.
3. Watch the direction of travel, not the level. The useful signal in a recombination is which old craft is being asked for. That is visible in job adverts months before any classification catches it, and years before it reaches a dataset like this one.

Somebody has to do the work of naming new work, and classifications are slow on purpose. In the meantime the honest thing a site like this can do is keep the two kinds of page separate, and be clear that one of them is built on a source and the other on somebody reading job adverts carefully.