The way AI-generated images works is by using one of the image present in their dataset, then add noise to it until it become nothing but noise, before progressively removing it (denoising) in an attempt to recreate the image initially used, having forgotten it in the meantime, with the prompt submitted as the only indication. All this is, of course, is an invisibilized process, just as the initial image used is not known to us. And it's precisely this cluster of pure random noise that allows us to see this intermediate image, at once an image that is no longer one, and the potentiality of all upcoming images.