The occupation pages are a reference: you look something up. These are the other thing — an argument that cuts across every verified record we hold, about a question people actually ask.
One rule makes this different from an opinion page, and it is enforced by a test rather than by good intentions: every section must cite a verified record or a task judgement. A section with no citation fails the build. The citations are printed under each section and link straight to the record, with its source, its scope and who verified it.
There is no single number, and the honest reason is not that the data is missing — it is that almost every published figure counts something other than a job. Here is each one we could check, and what it actually counts.
Read →Not how hard the task is. Not how well-paid. The property that predicts whether a task moved is whether something other than a human being can say 'that is wrong, try again' — and say it a thousand times overnight without getting tired.
Read →Put a dozen skilled occupations side by side and the same shape appears. Machines have taken the parts of the work that produce material — the photographs, the drafts, the scripts, the routine changes — and they are fast at it. What they have not taken is the decision that the material is right, and in most of these jobs the law says that decision belongs to a named, qualified person who signs for it.
Read →We refuse to answer this with a number, and that refusal is enforced by a test that fails the build. This is the argument behind it: every ingredient a year-estimate would need is either unmeasured, reversible, or decided by someone who has not decided yet.
Read →Most writing about AI and work describes technology arriving and jobs changing afterwards. In the verified record the commonest sequence runs the other way: a rule creates a duty, the duty lands on a named role, and somebody's job description grows a paragraph before anything is deployed.
Read →Adoption is narrower than the coverage suggests and lands in the same three places. The interesting question is not which function buys the tool — it is which role inherits the work the tool creates, because in the verified record that role is almost never the one that chose it.
Read →Two of the five tasks went away almost completely, and they went first because they were the most specifiable. The two that did not go away got more expensive, because the pipeline multiplies whatever judgement you hold it to. The fifth task did not get automated — it got deleted, along with the person who used to brief it.
Read →The typing moved; the job did not disappear. In the one large agent-driven migration with published numbers, the human's messages went mostly to review, testing and CI, to challenging design decisions, and to pushing work to completion. The records around it say why that work does not shrink: faster output is not the same as correct output, and trust in the output is falling as use rises. What has changed most is the door in, not the job inside.
Read →
LO