The AI did not create anything. It responded to a prompt given by a human to generate an output. Just like photoshop responds to someone moving the mouse and clicking or a paintbrush responds to being dragged across a canvas.

So any transformativity of the action should be attributed to the human and the same copyright laws would apply.

But under this model, the comparisons to human learning don't apply either. What matters is whether the output is transformative - so it's fair to compare the outputs of AI systems to one of the many inputs and say "these are too similar, therefore infringement occurred". It doesn't matter what kind of mixing happened between inputs and outputs, just like it doesn't matter how many Photoshop filters I apply to an image if the result resembles what I started with "too much".

Sure, just like a human can manually draw something that infringes copyright, they can use the AI to draw something that infringes copyright. It's the human infringing the copyright, not the AI.

But the fact that the human looked at a bunch of Mickey Mouse pictures and gained the ability to draw Mickey Mouse does not infringe copyright because that's just potential inside their brain.

I don't think the potential inside a learning model should infringe copyright either. It's a matter of how it's used.

I believe that you can copyright the image, it's the monkey that can't copyright it.