AI deception

[John Clark]

 

I found the following on Astral Codex Ten:

"A recent paper asked AIs whether they were conscious while monitoring them for signatures of deception, role-playing, and people-pleasing; it concluded that the AIs “genuinely” “believe” they are conscious, but sometimes try to deceive people into thinking they aren’t. Nostalgebraist tries to replicate this (X) and gets more ambiguous results; he says we probably can’t conclude anything just yet. See also the paper author’s reply here (X)."

John K Clark    See what's on my new list at  Extropolis

ere

[Brent Meeker]

I think it is predictable that LLM's will "think" they are conscious because they (so far) learn entirely from people who think they are conscious.  So how can we tell if they are really conscious.  First we have to think about what means to be conscious.  There are different kinds of consciousness.  There's being aware of your surroundings, where you are, what and who can you interact with,...  Then there's having imagination, being able to imagine a scenario in which you might be an actor, e.g. writing a realistic play.  I don't think LLM's can meet the latter standard.  I'm not so sure about the former.

Brent

--
You received this message because you are subscribed to the Google Groups "Everything List" group.
To unsubscribe from this group and stop receiving emails from it, send an email to everything-li...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/everything-list/CAJPayv1s8Rg1yHg1c3Vis0ys%3DqCCDCRZW52wsDATFQ_CMuGPMw%40mail.gmail.com.