> what exactly stops AI from taking those new jobs too as soon as the need for them is recognized?

Sample efficiency, energy efficiency and training data availability.

The current tech stack is around 6 orders of magnitude behind human capabilities. We have scaled up a paradigm that's been built out continuously for decades (transistors, matrix multiplication). To close the gap requires completely reinventing the entire stack (e.g. spintronics, physical graph traversal). The most advanced tech that is only theorized and not yet even proven is still 4 orders of magnitude less energy efficient than a human, and we have no idea how to get close to the same sample efficiency.

The question you are fretting won't matter for decades, and by then our culture, social and economic framework could be totally different than it is today.

Rather than worry about something so abstract and misunderstood, I would suggest focusing your time and attention on what is within your control here and now: who do you want to help, and how can you be of help, such that your efforts are sustainable and rewarding.

And if you really want to fret something at the edge of your control, fret about how AI is being used today to manipulate you and those you know into believing things that aren't in your best interest but are in someone else's best interest. If you aren't paying, you're the product. Most free media exists manipulate you, and if it has been designed to be addictive then it's virtually guaranteed.