When LLMs make stuff up, call it 'confabulating', not 'hallucinating'

Similar suggestions have been made over and over, and they've never stuck. "Hallucinating" has already hit mainstream media. It's probably going to stay that way out of sheer inertia because nobody outside of a small group of nerds on hackernews is going to sit down and think about whether "hallucinating" or "confabulating" more accurately describes the nature of the error. The existing term already captures the general idea well enough.

When educating the general public about the risks and limitation of LLMs, I think "hallucinating" is a useful term - it is something people can understand and it conveys the idea of LLMs being somewhat random and unreliable in their responses. I'm not sure "confabulating" is so easily understood or accessible.

I agree that confabulate is more accurate and descriptive. But in general usage "halucinate" is a commonly understood word and "confabulate " is not. Using the latter runs the risk of sounding off-puttingly technical or obscure.

It’s too late. People already call it hallucination and confabulating sounds weird. The idea that you think you can reverse the trend with a HN post is interesting though.

I don't see this as a more suitable term; I had to check the definition of "confabulate" and the first hit in the macOS dictionary is "formal; engage in conversation" and the second is "psychiatry; fabricate imaginary experiences".

The fact I had to look it up to make sure, and that it isn't the primary definition, makes this a bad alternative to a well understood word which has already become established.

In terms of communicating with laypeople, I don't think this is a hill worth dying on. "Confabulating" is not a term I think that many people outside of the field of psychology are familiar with in the first place.

If you do want to use another term for laypeople, I think "bullshitting" or "BSing" would have connotations that are more relevant than "hallucinating".

I find the easiest way of explaining LLMs to laypeople is "Bulls*t Engine". If tuned well they're going to answer like a salesperson or internet troll: if they don't know they will BS before they don't answer. Its not hallucinating, or confabulation its BS.

That's not to say they're not useful. The ability to BS is well regarded among humans as long as you, as a consumer, have a decent BS detector. And like a good BS artist, if you stay within their area of expertise they can be really useful. Its when you ask them something that they should or almost know that they start to be full of s**.

The problem with describing factual errors as "hallucinating" is that they're no more or less hallucinations than when it generate correct content. The entire point of these LLMs is that it synthesises content not in the original input. It's always "hallucinating" - its just that sometimes it gets it right and sometimes it gets it wrong.

This is a better term and a suggestion I’ve already come across. If popularizers of science and tech as well as the AI space were to begin using the term consistently, the mainstream media would follow suit eventually.

> LLMs have no sensors to experience the world with (other than their text input) and (probably) don’t even have a subjective experience in the same way humans do.

Computers are not subjects, period, and to attribute intentionality to them is nothing short of projection and superstition. LLMs do not somehow transcend what a computer is. Computers are machines that transform conventional representations into other conventional representations, but these representations require human interpretation to have meaning (during which the representation is related to the conventionally assigned meaning which is itself the terminus that requires no further interpretation). That meaning or content is the intentionality in question. Your won’t find it in the machine, by definition.

A source of popular misunderstanding probably comes from the confusion of what the computer is objectively doing with the metaphorical language being used to talk about it. Even “confabulation” is, strictly speaking, metaphorical or analogical, but within the analogical schema we’re using, it is a better fit than hallucination.

The purpose of words and language in general is to communicate which includes intent and meaning. If I go ask 100 random people what confabulate means, I doubt even a third know what it means. But regardless of “accuracy” if I tell someone a robot hallucinated, they more or less get what I mean.

Besides both the general public and academics have been calling it “hallucination” for well over a year now. I’m sorry to say this ship has long sailed.

Semantics, but useful ones. By the way, since you wrote '(probably) don’t even have a subjective experience in the same way humans do.', I want to talk about that.

Couldn't llms (or rather generative transformers) sort of confabulate their own subjective experience? If you added other senses/input, just text, that could all act on neurons, a 'memory', a feeling of passing time (internal clock or whatever, this is weirdly a sense every viable human get, as do all complex animal lifeforms). Is this a technical possibility? If a GPT was able to remember and 'internalize' recent input (I mean it passed recent input in its training function, not at all like ChatGPT 'remember' previous input/output that he experienced very recently) and it's output, and you asked it 'why was that your response', would it confabulate too? Would it trick itself? And could it experience 'déjà vu' if the 'internalization' bugged?

The first time I heard somebody say an LLM hallucinated, I knew exactly what they meant, so I think hallucination is a good term.

> A more suitable term would be confabulation, which is what humans do when due to a memory error (eg. due to Korsakoff syndrome) we produce distorted memories of oneself. This may sometimes sound very believable to outsiders; the comparison with LLMs making stuff up is rather apt!

Why? LLMs don't have memories any more than they have sensors.

My dad has Parkinson's and one of the pamphlets said such neural diseases can result in confabulating behavior. So I can definitely see the relevance of confabulate instead of hallucinate. It turns out, some medical experts wrote these opinions recently, but I don't have access to the full text:

"ChatGPT: these are not hallucinations – they’re fabrications and falsifications" https://www.nature.com/articles/s41537-023-00379-4#author-in...

"Chatbot Confabulations Are Not Hallucinations" https://jamanetwork.com/journals/jamainternalmedicine/articl...

As a neophyte wordsmith I like discussions of word choice, especially on HN (disappointed to see this flagged).

While I agree in principle, I'm not sure 'confabulating' is doing enough work. My hypothesis is given two words--one more accurate and one less accurate to describe the same phenomena--people will choose the more expansive or imaginative notion. That is, people will choose a word which generates more ideas.

We are in a time of expanding imagination about what LLMs and "AI" will do in the future. By definition 'hallucinating' LLMs are ridiculous, but the thoughts extending from 'hallucinating' label are nevertheless expansive in a similar way to our general overall feelings towards these technologies. Whereas 'confabulation' does nothing for my imagination.

I had literally never heard this word before people started complaining about LLMs hallucinating. I think I'll stick with the one that makes sense and wait for dictionaries to catch up. Words literally can mean multiple things.

On a related note, are there promising models / techniques to detect these sort of instances?

Say for instance I summarize something and want to check that the result doesn't contain hallucinations (confabulations :)) or more specifically that the summary contains only information present in original text. What's current state of the art for something like this and how well does it perform? I've read some about entailment models and fine tuned LLMs for this sort of thing but haven't found many great resources.

All seem valid to me. The inputs to training are the "sensor" through which LLM "experience the world". Maybe interpolation or extrapolation would be acceptable too.

That ship has sailed.

https://en.wikipedia.org/wiki/Semantic_drift

Confabulate is definitely a more suitable and sensible term for this. It’s clearly semantically more descriptive and would do a better job communicating what is actually happening.

The thing is, “confabulating” will never stick. People like myself will enter discussions about it and insist that it won’t stick, and because of that, I certainly won’t start using that term and hopefully I’ll convince a few others not to either.

While I agree with you that it's a better fit semantically, it won't happen, but not for the reasons given in the thread. The real reason is not the ignorance of the general public, but -- IMO -- the fact that even highly intelligent AI/ML engineers themselves have a limited vocabulary* and generally do not know this word.

* this is the prevailing secular trend, at least in the US

This is what I've been calling it as well. Confabulation is much more appropriate as a term.

Why change something that is already a well defined term in the context of generative AIs ?

Neither really seem appropriate; they're both far too anthropomorphising. The machine isn't making an error of _thought_; it's simply accidentally incorrect instead of accidentally correct.

A similar point was discussed 12 days ago (233 comments): https://news.ycombinator.com/item?id=37773918

thank you for the description. as a non-native english speaker i haven't heard the word 'confabulate' before, while hallucinate i learned from a young age.

Unfortunately, people are calling script kiddies "hackers" and LLMs "AI", so there's no likelihood of course correction.

Why not simply say "autocompleted"?

There’s a book called “Because Internet” you should probably read.

Tldr; language changes over time.

The idea of “precise language, and that someone is “using words wrong” is not correct, and history shows, repeatedly, obsessively trying to enforce “correctness” in language doesn’t work.

So… this is no different from telling people not to use “like, someone told me…” in sentences.

It won’t stick, and perusing it is probably meaningless.

If people call it hallucinations; that’s what they are.

That’s what language is.

There’s some deep irony in talking about this with language models.

You're right but too late. The die is cast. The world is not going to change to suit you.

lol, nerds.

seriously though, hallucinating makes sense to me because it genuinely feels like it's seeing things that doesn't exist.

for example, it comes up with non-existent postgresql functions.

that's hallucinations right there, i sometimes wonder what gpt-3.5 is smoking - i wanna try it.

Given that the "I" in "AI" is already a strong misnomer, why not just go with the lies people get told and make a fast buck?

Too late, the ship has sailed.

who asked?

Go back to Lesswrong.