They're trained on a lot of images and text. The big ones are trained on terabytes. The prompts I read in the paper involved well-known concepts, too. These probably repeated in tons of training samples, too.

It's likely they had data memorized.