The problem is not the algorithm. The problem is the way they're trained. If I made a dataset from sources whose copyright holders exercise their IP rights and then train an LLM on it, I'd probably go to jail or just kill myself (or default on my debts to the holders) if they sue for damages.