Tasks are not jobs

Two numbers travelled together recently and pulled in opposite directions. The Center for AI Safety's Remote Labor Index, which tests AI models against real freelance work by comparing their output to a paid professional's own deliverable, reported that frontier performance more than quadrupled in under eight months. Over the same period, tech and finance were losing an average of 28,000 jobs a month, even as the wider economy added 113,000 jobs a month. Read together, the two numbers look like evidence for whichever story you already believed. Read carefully, neither one answers the question people keep trying to make it answer.

A task is not a job

OpenAI's chief economist has made the distinction plainly: exposure of a task to AI does not by itself mean a worker will be replaced. Jobs are bundles of tasks, and the bundle can be reorganised around a new tool without the job disappearing. The right question is not whether a task is exposed. It is what happens to the rest of the bundle once that task changes, which is a question about how an occupation reorganises rather than a single number a benchmark can answer.

He has offered a personal illustration of the same point: his own father, an economist, moved from mainframe punch cards to running regressions on a personal computer. The computer exposed a large part of his father's work to automation. It did not replace him. It changed what the job consisted of and made him more productive at what remained. Software development tells a version of the same story at larger scale: it is among the occupations most exposed to AI on paper, and employment in it has not contracted anywhere near the degree many forecasts implied. Both are single examples, not proof of a general law, but they are exactly the kind of evidence that should make anyone pause before reading a task-exposure figure as a jobs figure.

What the sector numbers actually show

The tech and finance hiring numbers are real, and they should not be explained away. But they should also not be read as a clean AI story. Firms losing headcount at a moment when AI adoption is rising is consistent with AI-driven displacement. It is equally consistent with companies using AI as the explanation for cost cutting that was already overdue after several years of overhiring, which is a much less dramatic and much harder to headline story. Both explanations can be true in different companies at the same time, and the aggregate number cannot tell you which one you are looking at. Treating a sector-level headcount figure as settled evidence of an AI jobs apocalypse skips past that uncertainty rather than resolving it.

What the capability number actually shows

The Remote Labor Index tests economically valuable freelance tasks across design, media, data analysis and web programming, and counts a result as a success only when its quality would satisfy a paying client, judged against a real professional's own deliverable for the same brief. On that measure, quadrupling in under eight months is a genuinely fast rate of improvement. It is also a rate of improvement measured from a low base: the same benchmark still put the leading model's success rate at a little over one in six of the tasks tested, meaning roughly five out of six pieces of professional-quality freelance work still failed. Both readings sit inside the same number. A rising trend line tells you where capability is heading. It does not tell you that the job built from those tasks has already been automated, because most of what the job actually required still was not delivered.

Two readings, held at once

One honest reading of all this is that digital freelance work faces accelerating pressure as success rates climb from a standing start. The opposing honest reading is that humans are still needed on the large majority of tested work, bringing background knowledge and real-world complexity that current systems still lack. Both readings are supported by the same evidence, and choosing between them by instinct rather than by looking at what is actually being asked of the work is how a benchmark quietly turns into a forecast it was never built to support.

The useful response to that uncertainty is not to pick a side. It is to stop asking whether a job will be automated and start asking which tasks inside it are moving, what that leaves for the person doing it, and what the occupation looks like once the bundle is rearranged. That is a harder question than either an apocalypse story or a boom story, and it is the only one with an answer worth acting on.

What that means for the people involved

Short-run displacement in specific roles is real even where the long-run picture is one of reorganisation rather than replacement, and it lands unevenly: people who can move to where the bundle is shifting are affected differently from people whose specific skill was most of what they had to offer. Treating every exposed task as a countdown to a vanished job serves nobody well, and neither does pretending the disruption is evenly distributed or painless for the people living through it while the occupation-level story works itself out.