Except, LLMs don’t cite their sources accurately (if at all). A Google search may yield information with a source I can verify; my understanding is that LLMs are fundamentally incapable of doing so.
An LLM backed by an embedding store absolutely backs its sources. It’s a solved problem. But people cling to the narrative because that takes the fear away.
If one gave you a 10,000 page bibliography would you read all the sources? The same problem exists (long bibliographies) in human knowledge transfer. I’m suggesting this is an architectural issue, not a LLM problem per se. LLMs are not supposed to be AGI.
An actual citation and a string that happens to be formatted to look like a citation are very different. I wish we were at the point that most LLMs could do the former.
I've seen the former happen for some GitHub docs and when enabling plugins.
I agree there's cause for concern, and I hope it's a gradual rollout that's introspected in between each propagation of such, but I definitely think it's viable to be used no more less safely than other tooling in the government today (for better or worse).