A properly trained AI should not be able to do that. The data AI stores is not images or fragments of images. It’s a set of weights for various attributes for each term. Like the concept of “cat” would be stored as the set of most common values for various attributes it analyzed and found to match on all training images labeled as “cat”. With thousands of cat images as input, having proper variations between them, the result will always be unique.

It’s the same as a child learning to draw. If they see a drawing of a cat, they might try to copy that as best they can. But if they see many different representations of a cat then they will also learn to express themselves creatively and make up their own variation. And nobody is going to sue the kid for having looked at copyrighted pictures of cats.