With the current length of copyright, there is no possible way to build it otherwise. Saying that AI needs to only train on public-___domain works, is equivalent to saying that it shouldn't exist.
Of course, if it didn't exist, I wouldn't be buying art from artists instead. I'd simply go back to my stories not being illustrated at all. So from my perspective, this story is all about artists wanting to tear the ability out of my hands.
One could procure the necessary rights from the copyright owners / artists for all the training data. Sure, it would be quite costly and take a long time - but it is possible.
No, it isn't possible. Tracing copyrights for many things is literally impossible. Might be owned by defunct company X that died out several decades ago and had records destroyed in a fire.
One does not need every single copyrighted object in the world, just a large, diverse and representative set. One would have to skip the cases where rights cannot be secured. This is possible, see for example Spotify.
Nonsense, the claim made by those that say the end of knowledge work is nigh is that an AGI is coming that is superior to human intelligence. Humans do not need to train and read / consume gigabytes and terabytes of past solutions to produce new work and they do it at a trivial fraction of the amount of watts.
Of course, if it didn't exist, I wouldn't be buying art from artists instead. I'd simply go back to my stories not being illustrated at all. So from my perspective, this story is all about artists wanting to tear the ability out of my hands.