Not sure if writing the prompt is the human element because an AI can easily write prompts. I think a stronger human element is in the training data (while that training data isn't dominated by AI generated content itself).

> AI can easily write prompts

All our current AIs need prompts to produce content. So, you'd need to enter a prompt so that AI can produce a prompt for you.

But I guess you could develop some variant which just generates prompts in a cycle. But most of the generated prompts will pretty bad, so it will in turn produce heaps of crap.

That's where the human element is still needed - try a prompt, evaluate the result, tweak the prompt again. If this doesn't produce a good result, try a different idea. It's not even that different from how artists produce content these days (the create <-> evaluation cycle). Current AI is not good enough to be able to judge the produced content (that's why it produces so much crap after all).

There's way more human input in the training data than in the prompt is my point, many orders of magnitude more. Of course there needs to be something to start some chain, but that's like saying a human needs to press "deploy" or "start", it applies to any tool. Even a perpetual motion machine needs someone to start it.

And your brain also has way more training data (all the knowledge you gained, all the art you've seen) than what you yourself produce as "art".

Not really unless you have perfect memory. What people do is learning rules and breaking them. The rules exist outside of your mind, you're just trying to conform in some ways and distort them in others. You do not blindly copy what came before. When drawing a portrait, there are the rules of anatomy, perspective, colors and light, and your medium of choice. A style is a particular combination that you know works, but you still have to know the rules in the first place. You study masters to learn what is and what is not important in those rules, not to recreate their works in details. I've not heard of any art classes that train you by copying everything that has been produced.

Generative AIs don't remember the original media, they just detect/extract patterns out of them. They aren't able to recreate pixel perfect anything. Ask it to give you a Mona Lisa and the result would just kinda resemble the original. (kinda what a mediocre artist would be able to produce going by their memory)

> I've not heard of any art classes that train you by copying everything that has been produced.

You don't have to learn it, because it's the second nature for humans. We learn by imitation. Babies learn to talk by imitating sounds from parents. Artists learn by imitating style by masters. That's what generative AI does as well.