The headline is a bit clickbait.. As a purist, I say this isn't "AI", it's "Machine Learning".. sophisticated deep networks trained on millions of public domain images that can retrieve and combine components to make new conglomerate images.
I've been playing with MidJourney. I don't think it has any actual understanding of what it's doing.. for example, given the prompt "bruce campbell", you get this:
An impressive likeness for sure, But some of these look almost like exact stills from Promo material from some of his work. Similarly, given my name "randall hand" it seems to have been trained on some of my public imagery as well. The image below took a bit of triage, but it's the result for "randall+hand profile photo photorealistic"
I would love to know more about what data MidJourney is using for their training set. I would bet they're absorbing a few public domain datasets of public imagery and classical art, and then probably a mountain of data from sources like InstaGram. This would give them image itself, and the textual description provided by users (same as their current prompt system).