Conversation

AI art works by generating matrices that can be used in multiple contexts. These matrices will generate images that are close matches to features of large portions of the training data. This is why the images often look good until you look more closely or think about the meaning or the relationship between elements.

It's sort of like learning to paint by muscle memory and technique alone. If a line curves like this it keeps going, if its fuzzy textured it's part of a cat.