For those who wonder why all our generations didn't generate the dreaded signatures, I doubt it's because the training data was pruned of signed works. It is trained with public domain art ; il you can make Rembrandt-style, you would get a signature as well. Also, the model is obviously able to do NSFW art, so it was trained on it. I think they trained the model on a very broad basis. On the other hand, it's probable they add a better captioning system. OpenAI released several products in the same time window: Dall-E 3 and chat-gtp "vision", which is an extension that let ChatGPT analyse pictures.
My guess is that they developped the latter first and used it to make better captions of the images they used to train. The reason signatures appeared in previous AI systems is that the captionning was often very basic. For example, a picture of "a dog". Not "a dog on a table". Generally, it's OK because the AI has ample opportunity to learn table and differentiate it from the dog. But for "rarer" ideas, even if the image basis is large, it is possible that signature were never mentionned. And if you're learning a handful of pictures of say, an remorhaz, and all of them contains a signature, how would the AI be able to differentiate the remhoraz in the center of the picture and the thingie in the lower right of the picture? So when you ask for a remhoraz later, there was a probability to get... a signature. With better captionning, the risk of it happening is lower because everything on the image is designated.
Also, I doubt there is a post-image validation (it would take a lot of time and computing power, but maybe Microsoft has invested a lot into this tool), but it could be a reason we don't see any, if it flags images that contains signatures...
