Had a thought... If an LLM could weigh the credibility of sources of input, would that help with truth/hallucinations? It seems like the problem is they pull in input indiscriminately.
I don't see a good way to determine 'credibility' in an automated way. What I think I'm talking about is an opinionated LLM. Which I don't necessarily have a problem with. Sarcasm detection could be difficult.
I feel like this is a thing that should be obvious to people closer to the problem. Hm.
@toba and an opinionated LLM could not hold even a pretense of neutrality. Do you grant credibility to the NYT? Fox? The Intercept?
Which, frankly, as I think about it more, is probably unavoidable. If they did manage to address that issue, by whatever approach, It would reflect the biases of its creators in an ever more transparent way.
@toba yeah I've long thought that the central problem of all AI is the explainabilty problem. Even if they gave perfect results, as designed they wouldn't be able to explain their 'reasoning,' really. If that was how they produced them.
Though that does apply to humans, too. We often deceive ourselves, making decisions and then rationalizing them after the fact.
John Stoner (he/him)
Eric Stein �