It doesn't matter, even choosing a picture is a creative act that imparts copyright. Maybe if you hook up GPT-3 to generate prompts for DALL-E and then filter the images with another automated method, then the work is completely AI generated and deserves no copyright.
As for the authors of the training data, they still enjoy the copyrights over the works they created. Images generated by DALL-E only copy the style, something humans are also allowed to do. Anyway, you can't put the toothpaste back in the tube, now we have to live with image and text models. Next year we'll be shocked at video models that can play games, accomplish meaningful office work and also generate long videos.
As for the authors of the training data, they still enjoy the copyrights over the works they created. Images generated by DALL-E only copy the style, something humans are also allowed to do. Anyway, you can't put the toothpaste back in the tube, now we have to live with image and text models. Next year we'll be shocked at video models that can play games, accomplish meaningful office work and also generate long videos.