AI art isn't bad because of its inherent quality (though tons of it is poor quality), it's bad because it both lacks the essential qualities that people appreciate about art, and because of the ethics around the companies and the models that they're making (as well as the attitude of some of the people who use it).

AI has no concept of the technical concepts behind art, which is a skill people appreciate in terms of "quality," and it lacks "intent." Art is made for the fun of it, but also with an intrinsic purpose that AI can't replicate. AI is just a fancy version of a meme template. To quote Bennett Foddy:

For years now, people have been predicting that games would soon be made out of prefabricated objects, bought in a store and assembled into a world. And for the most part that hasn't happened, because the objects in the store are trash. I don't mean that they look bad or that they're badly made, although a lot of them are - I mean that they're trash in the way that food becomes trash as soon as you put it in a sink. Things are made to be consumed and used in a certain context, and once the moment is gone, they transform into garbage.

Adam Savage had a good comment on AI in one of his videos where he said something like "I have no interest in AI art because when I look at a piece of art, I care about the creator's intent, the effort that they put into the piece, and what they wanted to say. And when I look at AI, I see none of that. I'm sure that one day, some college film student will make something amazing with AI, and Hollywood will regurgitate it until it's trash."

But that's outside the context of your original post, in which you said that AI art would someday be better than what humans can make. And this is where my point about video game graphics comes in. AI is replicating the art in its training set, much like computer graphics seeking realism are attempting to replicate the real world. There's no way to surpass this limit with the technology that powers these LLMs, and the closer they get to perfectly mimicking their data and removing the errors that are so common to AI (like the six fingers, strange melty lines, lack of clear light sources, 60% accuracy rate with AI like ChatGPT, etc.), the more their power requirements will increase and the more incremental the advancements will become. We're in the early days of AI, and the advancements are rapid and large, but that will slow down and the hardware requirements and data requirements are already on a massive scale to the tune of the entirety of the internet for ChatGPT and its competitors.