I'm not the guy but I guess it's that it takes a bit of skill to tell it what kind of pizza you want and to verify it's actually the right thing that you receive etc. So in this example the order should be pretty elaborate or there should be several panels where the waiter brings the wrong or not quite right pizza.

While telling AI what you need is extremely easily in one regard (it's natural language) it makes it often pretty difficult to be accurate at the same time. Also by the nature of LLMs the results are hard or impossible to predict.

However, I think it's a bit like printing something with a 3D printer and then saying "I made this". The 3D printer actually made it but telling it what you want was the difficult part and at some point involved some 3D modelling or CAD or even g-code programming, tinkering with filament choices, speed and temperature settings, infill, support structures etc etc. While this is quite easy to do nowadays I remember the time where it was a big challenge to even get the damn filament to stick to the build surface. Another similarly is that it's often not the right tool for a job. E.g. if you want the same object thousands of times or objects that have super fine structures or objects that have to withstand a lot of physical use / abuse or temperature.

However, unlike with a 3D printer you're pretty much guaranteed to get something when you use LLMs.