[papercrane]

It's when the models outputs are connected to sources.

So for example, you might have a company chatbot that responds with a summary of a Confluence page it was trained on and a link to the source.

Doesn't eliminate hallucinations, but it does reduce them and gives you a chance to verify.

[ColdStream]

That works ok until the citation material ends up with LLM output being fed in. Already there are many examples of this happening.

[papercrane]

Yes, the trick is having a "ground-truth" KB it can link to, and trying to avoid citogensis from happening. I'm not sure if it's a technique that's every going to scale because of that.

[stingraycharles]

I don’t think this is correct, as we’re talking about all online discourse being influenced by AI, and you don’t know what to trust anymore, rather than generating AI output yourself (which is what you are referring to).

[papercrane]

In this context, it would mean building a knowledge base the model can link to that is considered authoritative. So maybe the LLM is has some slop in it's training, and that causes it to try and output some junk, but if it can't match it to it's verified ground-truth KB it doesn't end up outputting it.

The problem is figuring out what is the authoritative sources.

[stingraycharles]

But we're talking about AI writing and/or influencing other people's online discourse, not?