[gwerbret]

A bit tangential and pedantic, but:

> At the heart of the problem is the tendency for AI language models to confabulate, which means they may confidently generate a false output that is stated as being factual.

"Confabulate" is precisely the correct term; I don't know how we ended up settling on "hallucinate".

[delecti]

The bigger problem is that, whichever term you choose (confabulate or hallucinate), that's what they're always doing. When they produce a factually correct answer, that's just as much of a random fabrication based on training data as when they're factually incorrect. Either of those terms falsely implies that they "know" the answer when they get it right, but "confabulate" is worse because there isn't "gaps in their memory", they're just always making things up.

[tim333]

The Wikipedia version of confabulate:

>production of fabricated, distorted, or misinterpreted memories

doesn't fit. LLMs don't produce distorted memories, they guess random stuff and then forget it shortly after.

I think 'guess' may be more accurate.

[neRok]

About 2 years ago I was using Whisper AI locally to translate some videos, and "hallucinations" is definitely the right phrase for some of its output! So just like you might expect from a stereotypical schizo: it would stay on-task for a while, but then start ranting about random things, or "hearing things", etc.

[adammarples]

A hallucination is something that is experienced internally. Confabulation is better because it's the act of telling a porkie externally.

[torben-friis]

>fabricate imaginary experiences as compensation for loss of memory

Uh, TIL. This is wildly different to the Spanish meaning, confabular means to plot something bad (as in a conspiracy).

Which is a weird evolution in both languages, as the Latin root seems to mean simply “talking together”.

[rsynnott]

I mean, neither is a great term, in that they both refer to largely dissimilar psychological phenomena, but confabulate is at least a lot _closer_.