- generate details and texture, which photobashing was already used for; but that mostly solve the licensing problem for it
- generate random inspiration boards, for which image search was used (again, mostly solve the licensing problem)
- generate derivative stuff, e.g. typical game portraits or props
In practice, only the third point is generally discussed, because it lowers tremendously the entry barrier to generate images for people without any skills. It's like if you could pick up screenshots from other content, clean them up, and you're free to use it.
Whereas it essentially does not work for:
- cartoony generation. It relies too much on line consistency, visual clarity and abstraction
- concept design --not the flashy 10 minutes speedpaint type, but where you have to combine ideas in meaningful ways. In particular hard-surface design which required good 3D thinking and consistency of the whole.
These fundamental flaws are omnipresent, but can be hidden by certain styles where things are implied by color blobs, hidden by stylized brushstrokes, or simply an overflow of details (something Midjourney is very good at).
All in all, it feels like AI is a danger for people at the bottom of the profession hierarchy, but will elevate people at the top, whose work cannot be replaced. In other words, people who are more akin to be considered "artisans" rather than artists, who will take a prompt and simply clean it.
In particular, drawing has something like the 20/80 rule, where all the creative input is in the first 20% and the rest is 'rendering', a very mechanical task which you can mostly do with your brain turned off. As Yumenoley put it, "it was a mistake to let the AI do the interesting part".