What was said
The standard replacement prediction takes a job exactly as it stands today — the nurse, the teacher, the support agent, the junior developer — and asks whether a model can do that. Timelines are then attached to the model's capability curve, as if the job were a constant and only the machine were changing.
What actually happened
In case after case, the job that got automated was not the job as it was practiced a generation earlier. The sociologist Allison Pugh spent years documenting what she calls connective labor — the work of making another person feel seen, which is the load-bearing part of teaching, nursing, therapy, and much of management. Her history shows that for roughly forty years before the models arrived, management had been rewriting those jobs into scripts and metrics: service work got word-for-word scripts, professions got KPI dashboards, American physicians got electronic records that, by the counts she assembles, eat about two hours of clerical work for every hour spent with patients. Once a job has been turned into a script, a script is precisely what a model can learn. The automation did not conquer the job; it inherited a job that had already surrendered its unscriptable parts.
Engineering shows the same shape in a different uniform. The first tasks the new tools absorbed were not the senior engineer's judgment calls — they were the small fixes, the boilerplate endpoints, the test coverage nobody wanted to write: exactly the tasks the industry used to train juniors into engineers. The destination survived; the on-ramp was removed. In both stories, what disappeared first was not the work but the path that produced people able to do the work.
Why the predictions missed it
Forecasters model the wrong variable. They extrapolate the capability curve and hold the job fixed, because the job's transformation is slow, managerial, and produces no press releases. Nobody announces "we have finished hollowing out this profession"; it happens one SOP, one dashboard, one script at a time, over decades. A prediction that photographs the job at the moment the model arrives will credit the model with a replacement that management spent forty years preparing.
The unexpected part
What resists replacement turns out to be what resisted scripting first. Pugh cites a study of some 15,000 crisis-line conversations: callers in crisis did measurably better when counselors abandoned canned responses and answered creatively — the unscripted part was not decoration, it was the effective ingredient. Reading a room, going off script, exercising judgment under real constraints, carrying responsibility for the outcome: none of it survived being written into a manual, which is exactly why none of it has been automated yet. The frontier of automation is drawn less by what models can do than by what forty years of scripting failed to flatten. Before asking whether a model can replace a job, ask what the job had already been turned into. If the honest answer is "a script," the prediction was never about AI. It was about the script's authors.