Hacker News Reader

Comment by gwright | Hacker News Reader

That is an interesting point. I'm not sure why, but "hallucinate" doesn't bother me as much as "lie". Maybe it is because of all the ancillary baggage that "lie" has related to our current culture wars (lies, misinformation, disinformation, gaslighting, etc.)

I just asked (another anthropomorphic verb) ChatGPT for a better word and it came up with "fabricate".

> The term "fabricate" could be used to describe AI's tendency to generate fictitious information. This term has less anthropomorphic connotations than "hallucinate" and directly refers to the creation or construction of something, which could be a suitable representation for what an AI system does when it generates false data. Other alternatives could be "synthesize" or "generate," depending on the specific context.

View on HN

I always thought that "confabulate" would have been the best word to describe what LLMs do:

> Confabulation is distinguished from lying as there is no intent to deceive and the person is unaware the information is false. Although individuals can present blatantly false information, confabulation can also seem to be coherent, internally consistent, and relatively normal.

> Can include autobiographical and non-personal information, such as historical facts, fairy-tales, or other aspects of semantic memory.

> The account can be fantastic or coherent.

> Both the premise and the details of the account can be false.

> The account is usually drawn from the patient's memory of actual experiences, including past and current thoughts.

> The patient is unaware of the accounts' distortions or inappropriateness, and is not concerned when errors are pointed out.

> There is no hidden motivation behind the account.

https://en.wikipedia.org/wiki/Confabulation#Description

Confabulate is more of a neurological term so I think that's why it never caught on, but I've always thought that was the perfect word for what's happening

"Fabricate" has the same problem as "lie" and "hallucinate". All of those terms imply a cognition that isn't actually happening.

I think the best way to refer to these things is the more accurate "error". The LLM isn't lying, it's in error.

It doesn't. Hallucination in this context means 'an unfounded or mistaken impression or notion'.

There's one side of the current zeitgeist that over-anthropomorphizes these models. But there's another side that seems to be terrified that LLMs could be anything resembling intelligent

Both sides are pretty emotion over facts though. A lot of their "human-like" behavior is emergent from things that don't work the same way in a humans... but we also don't understand cognition enough to then say there's no overlap between what LLMs are doing and what some part of our own thought process is like. If anything we have more studies that imply the opposite going back decades: https://www.sciencedirect.com/science/article/abs/pii/S09266...

> It doesn't. Hallucination in this context means 'an unfounded or mistaken impression or notion'.

Correct. Which, as near as I can tell, isn't the correct characterization of what's happening with LLMs.

> But there's another side that seems to be terrified that LLMs could be anything resembling intelligent

I certainly have no such fear. I say that only to indicate that my biases are not of that sort.

The first sentence seems contradicted by the second no? If you don't have that fear then why bother playing around the semantics of "impression or notion"?

Is it an unfounded or mistaken cluster of tokens which resemble an impression or notion then?

At the end of the day it's producing an output which systemically was intended to be accurate, but ended up not being accurate through a mistake in generation: That's a hallucination.

I think "fabricate" is much less aligned with cognition.

It is completely reasonable to say that "This machine fabricates widgets." or "This algorithm fabricates pseudo-random song lyrics." without being worried that the machine or the algorithm is exercising "cognition".

Reply on news.ycombinator.com