> There is no IP theft because LLM outputs aren't protected, just egregious ToS violations

I meant original IP theft that occurs to train LLMs in the first place. But sure that implies that further LLMs based on that LLM are also tainted by that original IP theft.