It has been brought to my attention :) that there are at least two phrases on which Midjourney (v4) is so overtrained that it would be hard to argue that the outputs aren't derivative works: "afghan girl" (which produces slight variations of the justly-famous Steve McCurry photo of Sharbat Gula) and "starry night" (which produces somewhat more varied variations of the van Gogh De sterrennacht).

The Midjourney folks have actually banned "afghan girl" (and even "afgan girl") as prompts; seems likely that this is to help prevent their users from using their system to do something that might be violating copyright.

I wonder if there are other phrases that produce something so famously linked to them that they are instantly recognizable. And if so, how many are still in copyright.

This has somewhat softened my prior opinion that the training set is so large that it will basically never reproduce a single image from it. I do still think it's pretty rare, and only in extremely famous cases that occur a ridiculous number of times in the set. But still!

"The Mona Lisa" produces obvious (but much worse) variants, but that's definitely out of copyright...