My hypothesis:

One way to bias the generative output away from ugly generic sloppy defaults and towards something you personally prefer and desire based on your taste ... is to show it examples of what good looks like to you.

It is very hard for us to describe in words why you like a certain thing. But you can spot it when you see it. Or you can describe some aspects of it that you like or dislike after you see it -- but not before.

This approach is more accessible to casual creators -- who do not have formal training and experience to say "use corporate memphis" or "punk fanzine look with grunge typography".

(and this can apply to many fields -- not just visual design)