poetry is an outlier because most human poetry is meaningless pretentious dogshit that even GPT-3.5 could easily excel at generating. it's regular prose where their uncanny valley tone is immediately apparent to anyone who's been dealing with them for a while. all models are being trained to be bland, inoffensive sycophants, and you can't defeat their nature even with a 50k token system prompt.
But I'll take your point that poetry is different than regular prose. And I agree that even GPT-3.5 excelled at generating it.
I do think you can get a model to generate spicy, somewhat offensive/edgy content, but even then it doesn't sound like a real person. So I think bland/inoffensive isn't enough of an explanation.