What this is for

The short version

Most of what gets said about AI and jobs is either a sales pitch or a scare. Both are cheap to produce and neither helps the person asking the only question that matters to them: is this coming for my work, and what do I do about it?

This register answers that for all 830 occupations the US Bureau of Labor Statistics counts, one at a time, against a rubric published in advance, in language specific enough to be wrong. Then it shows its work, and lets the people who do the job argue with it.

What we're trying to do

Tell people the truth early enough to matter

A verdict that arrives after the layoffs is a headline, not a warning. Scoring capability exposure rather than adoption means the register can say a job is exposed while it still looks fine from the inside — which is uncomfortable, and is the whole point.

Give every answer a next step

A score on its own is just anxiety with a number attached. Every occupation page carries what's most at risk and what survives, which skills hold, and — where an honest one exists — where the experience transfers. 431 occupations currently carry a transfer route and 830 carry a set of conditions that would move the score up. Where nothing clears the bar, the page says so and explains why rather than inventing a hopeful answer.

Make the work findable under its real name

Almost nobody calls their job by its federal occupation title. The register is searchable by 45,617 job titles people actually use, so finding yourself doesn't require knowing which SOC code your employer files you under — and where that code doesn't mean what the title suggests, the page says so before you read the verdict.

Be the record, not the take

Scores are judgements. Deployments are facts. The tracker and the dispatch keep those apart on purpose: what a named organisation actually did, with a source you can open, is a different kind of claim from what we think a job's exposure is — and the difference between "could be automated" and "somebody automated it" is most of the argument.

What we promise — and how you can check

Every line below is a commitment with a mechanism behind it. If one of these is ever untrue, it should be provable, not arguable.

• The method is published before the verdicts. Five dimensions, 0–20 each, summed. The arithmetic is printed on every page so you can check the total yourself. The rubric.
• We publish our misses. The worst call in each direction is named on the methodology page and picked by the data, not by us — currently Hairdressers, Hairstylists, and Cosmetologists, scored 75/100 and SAFE, which went on to lose 21% of its workforce. A register that only reports its hits is a marketing document.
• Uncertainty is stated, not smoothed. Re-scoring the same occupation moves it a few points on its own, so every score carries its own re-score band, and verdicts near a boundary are labelled as near a boundary.
• Sourced claims or no claim. Every entry about a named company requires a working source URL; the build refuses entries without one. Sources that won't open to an ordinary reader are dropped rather than cited.
• Quotes are verified mechanically. Where a page quotes a federal definition, a validator checks it verbatim against the government's own file and fails the build on a paraphrase.
• You can argue, in public. Disputes are open, and every change to an occupation's score is kept with its reason, so the register can be audited rather than trusted.
• Scores move on evidence, not on a schedule. A monthly refresh would mostly republish our own noise. An occupation is re-examined when deployments accumulate against it, which is why a verdict changing here means something.

What this refuses to do

This matters more than the promises, because these are the ways a site like this normally goes wrong.

• No predictions with false precision. Nothing here says a job ends in a given year. Exposure is a judgement about what the work involves, not a forecast of what employers will do, and the two are constantly confused in public.
• No fear as a growth strategy. A COOKED verdict is the beginning of a page, not the end of one. If the honest answer is that an occupation is fine, the register says it's fine — 176 of them are.
• No selling the anxiety we create. Where a page points at training, the link is what it is. Nothing here is a paid placement dressed as advice.
• No pretending the numbers are ours when they aren't. Employment and wages are the BLS's, unmodified. Job titles are the Department of Labor's survey. The handful of titles we add by hand are shown separately, because borrowing a federal source's authority for our own judgement is the easiest lie to tell here.
• No advice we're not qualified to give. This is not legal, financial or career counselling, and it doesn't know your employer, your industry, or you. It compares whole occupations — that's genuinely useful, and it is also all it does.

Where this is weakest

Scores are written by a language model against a fixed rubric, which makes them consistent but not independent — and that model already knows how earlier automation waves turned out, so agreement with history is a consistency check rather than a successful prediction. The register has been tested against actual employment change and the relationship is real but modest: it explains under a tenth of the variation. Occupation-level scoring also can't see specialisation, industry or seniority, which is often exactly what decides whether an individual is exposed. All of this is written up, with the numbers, on the methodology page — and if any of it stops being true, that page changes first.