What if someone trains a model on public domain works only?

Also adobe sort-of did that already - they trained their own model only on their own images. So no "IP theft" there.