How we decide

The one rule

Jobs are not atoms. Tasks are. Nobody is replaced by AI — tasks are. So we never start from the job title. We split the job into 4–8 real tasks, judge each one separately, and only then roll them up into a headline verdict. If the headline ever contradicts the task breakdown below it, the task breakdown wins and the headline gets fixed.

The three verdicts

Verdict	What it means
SAFE	The core of this job is structurally resistant. AI becomes a tool, not a replacement.
SHRINKING	Significant parts are being automated. The role narrows and shifts — it does not vanish.
ON THE MENU	The core tasks are going, and what is left will not support as many people. A time horizon applies.

A verdict never stands alone. It always ships with the task breakdown, the resistance tags, and — for anything that is not safe — a year. If we cannot write those, we do not publish the entry.

Task-level verdicts

• gone — a competent practitioner today already delegates this to software. Not "could in theory": actually does.
• going — the machine does the first draft; the human reviews, corrects and owns it. Hours collapse, the task survives in reduced form.
• safe — blocked by a structural reason, not by a temporary capability gap. Every safe task must name at least one resistance tag.

Resistance tags

The question is never "is AI smart enough?" — it is "what structurally stops a machine from taking this, even if it is smart enough?" Capability gaps close. Structural walls do not, or close far more slowly. These are the walls we recognise:

Tag	The wall
physical-presence	Requires hands and a body in the physical world.
legal-liability	A human must legally own the outcome and sign for it.
regulated	A licence, permit or statutory wall stands in the way.
trust-relationship	The value is a personal trust relationship, not the output.
human-judgment	Contextual decisions under uncertainty that nobody wants to delegate.
creative-taste	Aesthetic judgment: AI can generate, it cannot choose.
accountability	"Who gets blamed when this is wrong" demands a person.
physical-context	You have to be on site, in the room, at that moment.
emotional-labor	The emotional labour is the job itself.

An entry carries 1–3 tags, strongest first. If the only argument for a job's survival is "AI is not good enough yet", that is not a tag — that is a countdown, and the entry gets a year instead.

The "safe until" year

This is the most provocative number on the site, so here is what it actually claims: the year by which we expect the core tasks of this job to be routinely machine-done in ordinary practice — not the year the job title disappears, and not the year the technology first becomes possible in a demo.

It accounts for three lags that people forget: capability has to arrive, organisations have to adopt it, and regulation has to allow it. That is why regulated professions get later dates than their raw task difficulty suggests.

It is an estimate, and it is meant to be argued with. That is the point.

What we refuse to do

• No doom for engagement. Every entry, including the red ones, ends with something you can use. If a verdict leaves you with nothing to do, the entry is incomplete.
• No hedging into meaninglessness. "It depends" is not a verdict. We commit, and we publish the reasoning so you can attack it.
• Sponsorship never touches a verdict. No sponsor has ever seen an entry before publication, and none ever will. The full rule is here.
• No fake precision. We do not put a percentage on a job. We do not have that number and neither does anyone else quoting one at you.

Sources and drafts

Entries reviewed against published labour data, regulatory rules or industry evidence carry a Sources section. Entries without one are marked community draft on the page — the argument may still be good, but nobody has attached evidence to it yet. Attaching that evidence is the single most useful contribution you can make.

How verdicts change

Every entry carries a lastReviewed date, shown at the top of the page. When a major model or product ships, we re-open the affected entries and move verdicts if the task breakdown actually changed — not because the news cycle is loud. A verdict moves when a specific task changes state, and the headline follows.

Every change is recorded and dated in the verdict changelog, because a site that quietly rewrites its own predictions is not worth reading. If you are citing a verdict, cite it with its review date — it will go stale, and that is by design.

Disagree

Each entry is a single JSON file in a public repository. If you do the job we wrote about and think we got it wrong, you are a better source than we are: open a pull request, change the verdict, and make the argument in the description. If it holds, it ships.

Contribute on GitHub · Browse every verdict