John Fechtel in the Hedgehog Review:
Even if it uses a fantastically complicated machine playing games with image combinations, any model is only the sum of what it is trained on. In the case of DALL·E 2, the model is trained on a comprehensive corpus of humanity’s collected visual output. Some hopeful proponents call it a kind of imaginative mirror: something that can reflect our own mental image back to us. We could think about our own visual imaginations in terms of models, too. They are trained on some smaller and less complete set of humankind’s vision. There is a potential gap, though, between our imagination and the computer’s “dreams about electric sheep.” Can this model cross that gap?
Spend any time tapping prompts into a generated art engine and you will inevitably discover a sense of the uncanny in the image responses. The nearest analogue I can think of is the slightly unsettling sense you have when someone isn’t quite looking you in the eye, an almost-but-not-quite connection of understanding that feels as if it’s running two or three parallel tracks to the left.