That's more an argument for why you can't just use LLMs as a source of truth. Conveniently, LLMs like ChatGPT do often cite their sources, especially if you prompt them to.

Maybe a nit: LLMs do not and cannot cite their sources (at least scraped sources for the purpose of training)

It’s kind of the harness that is doing the citing (or providing the context for the model to).

But an LLM sans search can reproduce some copyrighted work with minor variations and there’s no way to know exactly where it came from.