These days the LLMs feed the LLMs so you can model models unless you're excluding any public data from the last decade. You have to assume all public data based on users is tainted when used for training.
The truly valuable data is the stuff that was created prior to LLMs, anything after this is tainted by slop. Any verifiable human data would be worth more, which is why they are simultaneously trying to erode any and all privacy
They can produce high-quality answers now, but that's just because they wwre trained on things written by humans.
Any training on things produced by LLMs will just reproduce the same stuff, or even worse actually because it will include hallucinations.
For an AI to discover new things and truly innovate, or learn about existing products, the world, etc. it would need to do something entirely different than what LLMs are doing.