Both of the usual scripts are wrong
One script says everything is automated within five years. The other says this has all happened before and it always works out. Neither survives contact with what is actually going on.
Here is a frame that does better.
Models take tasks, not jobs
A job is a bundle of tasks plus responsibility for the outcome.
A radiologist reads images. That is one task in a job that also includes deciding what to do next, discussing it with a frightened patient, arguing with a colleague who disagrees, and carrying the consequence when it is wrong. In 2016 a prominent computer scientist said medical schools should stop training radiologists. Ten years later there are more of them, and image-reading software is used every day in hospitals, as a tool inside the job.
So the first question is not is my job at risk. It is which of my tasks are at risk, and what is left when they go.
Which tasks move first
Tasks tend to be automated early when they are all three of these:
Routine and repeated, so there are plenty of examples to train on. Text or image shaped, so a model can actually touch them. Cheap to get wrong, or checked by someone before they matter.
Drafting a standard reply. Summarising a document someone will read anyway. First-pass translation. Boilerplate code. Formatting and tagging. Routine transcription.
Tasks resist automation when they involve responsibility that has to sit with a person, physical work in messy unpredictable spaces, judgement with no clean right answer, or relationships where the point is that a human is doing it. A plumber in a 60-year-old building is far safer than a copywriter, and that reverses everything people assumed about which jobs computers would take.
The honest part about the past
The cheerful precedent is bank tellers. Cash machines arrived and teller numbers went up for two decades, because branches got cheaper to run so banks opened more of them. That is a true story and it is used to end conversations.
The rest of it: the job changed into selling products, the growth stopped, and teller numbers have fallen hard since 2010. Meanwhile the same decades hollowed out clerical and manufacturing work across many countries and the people affected largely did not become something better paid. The economy adjusts is true, and it is useless if you are the adjustment.
What is genuinely different this time is which tasks are exposed. Previous waves came for physical and routine clerical work. This one comes for language, images and code, which is where a lot of educated, well-paid work lives.
The thing actually worth worrying about
Not mass unemployment next year. The entry rungs.
Juniors have traditionally learned by doing exactly the tasks that move first: the first draft, the basic research, the simple ticket, the routine translation. If those are done by a model, the training ground for becoming senior disappears, while the demand for seniors continues.
Nobody has solved this. Firms are noticing it and mostly not acting on it. If you are early in a career, this is the real thing to think about, and the answer is probably to get to judgement faster than the previous generation had to: learn to check work rather than only produce it, and learn the part of your field that carries responsibility.
What to actually do
List your tasks. Mark the ones that are routine, language-shaped and low-stakes. Assume those change, and use the tools on them yourself rather than waiting to find out. Then invest in the rest: the judgement, the accountability, the relationships, the physical and situated parts. Those are the job now.
Before you move on