Of course in the retrieval (R in RAG) you can do attribution of the source material: you’re just doing a search and bringing up a literal excerpt of something from your database. You know exactly where you got it.
But then for generation (G) you hand that excerpt to the LLM, and the only reason the model can “understand” it is the vastly larger corpus of text you trained the model on, whose origin is, due to the very nature of LLMs, smeared out of existence. That is the controversial aspect that (AFAIK) has no technical solution at present.
He seems to imply that because attributing the R part is easy, attributing the G part should be too, but those are completely independent problems. Not to mention that you don’t have to do a retrieval step to generate things in the first place; the LLM alone can do that.
The part about doing a retrieval on the output to see if it’s similar to something else is at least technically possible, but he handwaves past the problem of what on earth you’re supposed to do if you find something. YouTube doesn’t do a great job (e.g., people getting copyright strikes from their own performances of public domain works) and it at least has an unambiguous set of things to search for.