If you took a random concept and explained it to a person they could using their existing knowledge set, draw it somewhat competently. That is because people are able to apply knowledge to make something new.

Theoretically it can, but it would involve meticulous and proper labeling of each training data. Currently most of the trained data are automatically labeled and they’re not descriptive/verbose enough. I believe the improvements from the latest version of DALL-E is due to OpenAI’s use of a more advanced image labeler.