On the left, the prompt is "photograph of a superhero called manwoman --ar 2:4"; on the right, the prompt is "photograph of a superhero called womanman --ar 2:4". The results are similar if you put a space between the permutations of "woman" and "man". Either way, you get only women.
Whenever you ask Midjourney to generate a superhero photo, there's a fair chance you're going to get one dressed like Superman. I doubt that whoever owns the original images that were used to train the model gave their permission (although I suppose it's possible) ; how does Midjourney have enough images routinely to create Superman costumes, then?
Here's how I think this works.
The training data come from sites that the users who uploaded the images have agreed (in the conditions of use that the user didn't read) can be used for such purposes. This means the data-gatherers can claim their training set is ethically sourced, and they're covered against accusations of copyright violation. However, the people who uploaded the images didn't actually own the originals. Sure, the social media platforms they uploaded them to made it clear (in the conditions of use that the user didn't read) that users aren't allowed to upload content belonging to someone else, but their users did so anyway. As a result, merely taking images from a site where image-use has been granted doesn't mean that all the images are fair game. This is a big problem for AI models that require vast data sets: they can obtain permission to use images from given sources, but the people selling them access to those images can't guarantee that they're not copyright someone else. They won't get far if they claim it's the users who are to blame: if you steal from a thief, you're still a thief.
I particularly like how the person in the first image is facing left with a body facing right.