An argument I heard, and adopted is that it’s never “just” cosmetic. Your enjoyment of the game is impacted by how you perceive your avatar. This is why fortnite skins sell so well to new players. It’s not just cosmetic to drop $20 on Cuddle Team Leader. It makes a user feel silly and increases enjoyment running around as an obvious pink mascot costume. It prolongs how long you play both by increased enjoyment, and sunk cost fallacy. In any game with cosmetics, purchases drive playtime.
Generally I think this has something to do with average age of model trained vs age of depicted person. Humans have large heads proportional to their bodies in youth and then averages out. On top of this another commenter mentioned art. So many anime/cartoon characters have larger heads to show the expressions and in turn show a youthful nature in these characters.
Another addition to this is, as another poster stated, the angles of these photos makes a larger head more apparent, probably because we associate larger heads with intelligence, and so poses have just developed to make that appearance.
Similar to “the average number of arms is less than 2 because there’s greater chances there will be people with zero or one arm, rather than an extra”. Your model will draw on its wealth of samples where the head is disproportionately larger, instead of proportionate - because it was probably fed more photos of hat age range and doesn’t know any different. You asking for AI to give you a person like Trump (78) will show that disparity in head size more than if you ask for Greta Thunberg (21)