1. Simplifications of reality(the actual artist training method would be traditional studies off life and photo reference followed by gradual reduction and symbolization to a style)
2. Symbolic meaning. Things like the style of eyes, clothing, etc are all meant to signal personality. This is stuff that current AI techniques don't really touch upon in any direct sense.
Since the ML method is built on interpolating off final results, it's going to lack in these qualities and produce something that is consistently an "average impression". Akin to asking the algorithm to generate mythical heroes by mashing up the various stories: you get a hero that is somehow the average of Icarus, Heracles and Achilles, which would be less of a character than the originals.