I think the detail that makes it tricky is that, in theory, an AI could just have a single neuron and simply output verbatim the image it was trained on. In which case we obviously wouldn't say that the image it outputs is a new work and freed from copyright.

So clearly there is some threshold of "transformativeness" which an AI must possess in order to be trained on one image and produce a different one that is considered a new work. Where exactly that threshold lies is difficult to define from a legal perspective.