As you say, most generative work is empty because it's just form. Human creators manipulate form to communicate content. There's a message, and art is judged as much on the sophistication of the message as on the form it's carried in.

Form without content can be an interesting technical experiment, but it's culturally massless and kind of boring. Not to mention wrong in a destabilising way - like someone speaking random syllables that make occasional sense, but don't form a coherent message.

So the most successful generative work cheats by automating processes of cultural reference - like style transfer - or by adding levels of formal complexity that would be very difficult to handle by hand, but can at least produce something that seems a little original and surprising.

The less successful work doesn't even get that far.

Automated software generation is going to have similar problems. You have to parse intent, and then balance it against practicality. Getting that right is so hard humans really struggle with it, so I can't see machines being better at it any time soon.