It's trained on a dataset that contains images together with text describing the image. Some images have been scraped from artstation, and if they were scraped from the "Trending" page, where well done images end up, it's included in the description. So by adding that to the prompt, you influence the image to be more similar to images that have been trending on artstation.