Hacker News new | ask | show | jobs
by barbariangrunge 1055 days ago
Stop calling them hallucinations. If we're going to anthropomorphize AIs, let's just call it bullshitting and lies. If we're not going to anthropomorphize AIs, then we need a different term
18 comments

> If we're going to anthropomorphize AIs, let's just call it bullshitting and lies.

why? "bullshitting and lies" suggests that the AI is intentionally being deceptive. "hallucinations" conveys the idea that the information is incorrect, but the AI perceives it to be correct, which is more in line with what is actually happening.

I'd go further and say that the AI doesn't even perceive it to be "correct". It's just saying these words are likely to follow those words.
That seems more technically accurate but less likely to catch on since it's a much less common word. I think my parents are a lot more likely to understand from context a news story that mentions that "lawyers relied on an AI that hallucinated court cases" than "lawyers relied on an AI that confabulated court cases".
I feel like "fabrication" is just sitting right there begging to be used.
Bullshitting and lies is what the humans selling the AI-powered services are doing. Hallucination, delusion and confabulation are what the AIs are doing (and some of the humans, too).
"Making shit up in order to fulfill some requirement" is the definition of lying, so whether it's a human or an AI, just making shit up in order to generate prompted output is flat out lying. Not "hallucinating". And the best part is that until LLM get valitidy checks baked in, even the things they get right are lies if presented with authority, because the LLM doesn't know whether it's true or not. In fact, the LLM doesn't know, full stop. It's still just a very well crafted autocomplete, and literally nothing more. So if we're going to anthropomorphise, call them what they'd be when humans do the same:

lies, and damned lies.

> "Making shit up in order to fulfill some requirement" is the definition of lying

I'd argue that there is an element of intent or agency involved. When a human makes things up intentionally or by choice, that is lying. When they do it unintentionally, that is not lying. It is usually called confabulation (or, honest lying - where the actor does not know they are not telling the truth). I don't think AIs/LLMs have agency or the ability to make things up intentionally. They are just doing what they are programmed to do and everything they produce looks the same to them. It is all true as far as the LLM is concerned. They might be confabulating, but I don't think they are lying.

The AI intentionally makes things up because that's literally the whole point of the LLM concept, both abstractly and concretely. WE MAKE IT LIE by conditioning it to lie. The "AI" part is not separate from the humans who made it, we are part of the system, and we made a computer that constantly and continuously lies in order to generate seemingly credible responses.

We made a computer that lies, all the time, about everything.

"...Why?"

<insert the shouting robot comic here>

Your argument is that any untruth is a lie.

Do you consider fiction authors to be liars?

No, my argument is that if you don't KNOW what truth is, everything you say is a lie.
If I make a claim based on prior knowledge and statistics I’ve learned over time, it’s not lying if it’s wrong. Lying has intent. Plenty of people say incorrect facts that they think are correct.

In second grade, my cousin talked a lot about flax farmers in South America, after learning about them in class. Turns out the lesson was on quinoa farmers, and he forgot the original produce and “hallucinated” the statistics about flax farmers instead. Technically the term is confabulation. Was he lying? No because he wasn’t trying to tell us fake facts.

LLMs have no intention of being wrong. Their “hallucinations” or whatever are just whatever makes sense from their statistical models. They’re really just confabulations.

"Bullshitting" seems like a good term for accurate or inaccurate responses.

Let's extend "LLMs have no intention of being wrong" to "LLMs have no inherent sense of being correct" - sometimes their predictions happen to be correct, sometimes they don't. But they're all hallucinations generated from the same process.

Bullshitting always requires a hidden intention to manipulate.
Nah, it can be just talking without much rigor or verification of fact or anything, often with a loose boundary between opinion and fact - the "to talk in an exaggerated or foolish manner" definition.

As in "my buddies and I were bullshitting about movies the other day."

ChatGPT definitely talks with an exaggerated manner confidence-wise.

Mislead is probably a more accurate term. Bullshit exists outside of the correct/incorrect binary and serves only to build narrative.
To be fair, if we are going to anthropomorphize it, bullshit and lies implies some sort of negative intent that I’m not sure the models have.

Bullshit is probably the closest, as people will bullshit for all sorts of reasons, but hallucinations is at least intent-neutral, which I think is the point.

A person can 100% believe in the lies they've been told, but that person is not hallucinating.

Take for example climate change deniers; apart from the corporations and the politicians that abuse scepticism to maintain their power and wealth, many of the most fervent deniers truly believe the nonsense they're saying.

Perhaps a more neutral term like "falsehoods" is applicable here.

I think calling it lying requires intent, I’d just say those people are wrong rather than lying
no not true - lazy, imperfect and damaged cognitive functions have similar results.
In classification problems there’s a useful term for something similar already — False Positives

   false positive (FP), Type I error
   A test result which wrongly indicates that a particular condition or attribute is present
https://en.m.wikipedia.org/wiki/Confusion_matrix

Edit — Though I’m not sure how well that fits for a LLM (it’s more a series of false positives at each step of prediction in the sequence).

In psychology, we've got a term which is almost 100% matching: confabulation. The only part which isn't correct is association with brain damage.

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

In psychology, confabulation is a memory error defined as the production of fabricated, distorted, or misinterpreted memories about oneself or the world. It is generally associated with certain types of brain damage (especially aneurysm in the anterior communicating artery) or a specific subset of dementias.

I think it's a good term for this but it also side steps the issue that this isn't a actually intelligence coming up with it. It's just machine noise
It doesn’t matter if it’s noise, it only needs to be useful. Confabulations are not only not useful, they’re actively harmful.
Call them confabulations.

"Confabulation refers to the production or creation of false or erroneous memories without the intent to deceive, sometimes called 'honest lying'"

"Confabulation is the creation of false memories in the absence of intentions of deception. Individuals who confabulate have no recognition that the information being relayed to others is fabricated. Confabulating individuals are not intentionally being deceptive and sincerely believe the information they are communicating to be genuine and accurate."

https://clinmedjournals.org/articles/ijnn/international-jour...

The words confabulation and lie have the same problem when applied to the current state of "AI": they embody are certain level of intent; in the former case, the implication is that there was no intent to deceive, while a lie is the opposite. Still, they both imply intent, and for as far as I know nobody has been able to conclusively demonstrate intent on the part of a chatbot.

Hallucination doesn't require intent.

I like "fabrication" because it correctly implies the data is being pulled from somewhere and assembled. It does not put a judgment on the initial data, nor on the assembled product.

It can take on a positive or negative meaning, depending on the context.

"ChatGPT fabricated an answer that was technically correct, but misleading," or "ChatGPT was able to fabricate an innovative solution that had eluded us."

Edit: And of course, everyone's favorite, "ChatGPT found guilty of fabricating case citations."

https://www.techspot.com/news/98860-chatgpt-found-guilty-fab...

Hallucinations are defined as “Perception of visual, auditory, tactile, olfactory, or gustatory experiences without an external stimulus and with a compelling sense of their reality, usually resulting from a mental disorder or as a response to a drug.”

That doesn’t sound like what AI/LLMs are doing, at all. There is no mental disorder or drugs causing then to output what we would consider to be false information. The machine is not perceiving anything without an external stimulus. Everything they generate is from the stimulus we have given it.

Given the euphemism "bug" substituting for "programming error" you'd be tempted to allow something similar for LLMs, but these are not errors, the output is by design.

There is no motive for truth, just the most likely output, even if the likeliness is low.

> There is no motive for truth

This also ignores the larger question that has been a known issue for at least 2,000 years: "Quid est veritas?"

You can substitute "accuracy" or "usefulness" for "truth".
It’s the adopted term. I don’t see why it HAS to be the absolute exact closest possible term to what it would be in a human or something.

It feels a bit like saying “stop calling it e-mail! It’s got nothing to do with real mail!”

Because people feel like they have nothing to add so instead of not adding anything they decide they have an issue with some minute detail that doesn't really matter and then start raising hell.
IMO this whole concept of "hallucinations" is a made up buzzword (in the context of AI) to distract from the fact that the companies who are writing/training these models know full well that what they spit out is just as likely bullshit as it is "correct".

Saying "we have no idea if it's going to spit out something accurate" doesn't sell.

"oh it's hallucinating, how cute" is an easier sell.

Then tell us what we should call these manifestations...

It's say to say stop calling it X, but then what are we supposed to call them?

A popular term in the LLM space is 'confabulation': https://community.openai.com/t/hallucination-vs-confabulatio...

It fits better than the alternatives I've seen proposed.

Malfunctions? Breaking? Errors? Poor reliability?
The problem I have with this is it seems too generic.

Shouldn't we try to categorize the types of errors at least somewhat?

Sure, and it's probably wise not to pick error categorizations that impute consciousness onto a thing that's 1) almost certainly not conscious, 2) behaves quite similarly to a conscious thing, and 3) might become conscious at some point in the near or distant future

Hallucinations are definitionally features of conscious experience. Pick a different word or make one up!

I guess to me hallucination doesn't necessarily imply consciousness.

"an experience involving the apparent perception of something not present"

I guess it depends on exactly how you define "experience" and "perception".

But yeah there are better words than hallucination that are even more generic and do work better.

We should call it Twitter.
Bullshitting has a goal, hallucinations are random, seems apt.
I'm not so concerned with that as I am with the fact that this isn't one. Article says

> they tend to make up fake information – errors called “hallucinations.”

Hallucinations are a certain kind of error. But what appears to have happened here is a _direct_ manipulation from Microsoft. Which is a risky play by them. It doesn't take much to erode trust. People tend to trust LLMs because they tend to get things right. But if people see a few things that they know is wrong, they will quickly stop trusting. If they see a few things as marketing, then they will very quickly stop trusting.

It's not a hallucination, it is a filter. Microsoft manipulated the output to prefer their own products and boy is that a risky strategy.

> Microsoft manipulated the output to prefer their own products and boy is that a risky strategy.

Makes me wonder how they plan to monetize these chatbots and if they won’t just fizzle out like voice assistants.

I don’t see how there won’t be concerns over asking a chatbot for the best pizza in town and receiving an answer like “Customers love the new Meat Lover’s Pizza from Pizza Hut! Brought to you by Pizza Hut… (list of pizza places here)”. Amazon couldn’t figure out how to make money off of Alexa, how are Chatbots any different.

Chatbots just seems like the marketing tool if we're being honest. I can't see any way to monetize them without destroying them. Even just having them could threaten their own existence (or make bloggers higher qualities if Google fucking decides to fix SEO...). LLMs on the other hand have plenty of use cases, though I think people are still way over hyped on them and aren't interested in how the boring stuff has good utility.
I wouldn't get this worked up about a simple term that keeps things understandable for a layperson, lest your head might explode once you see how people are anthropomorphizing some AI "companion" bots.
It's belief vs intent. Intent would be anthropomorphizing AI much more than belief and would denote a theory of mind. I'm not sure of a better term for things the model 'believes' to be true that are wrong. I think it's quite analogous given that the model then elaborates on the false belief in much the same way that humans appear to do with hallucinations.

Additionally belief does not mean human; for example animals can have beliefs, even very rudimentary animals. I think is more of a way of self-containing the entity and treating it as a black box.

I dont understand why you are so worked up about the term and i also dont understand how your characterization of it as bullshitting and lies is accurate in any way.
Ha! Yup, one of my friends who has been working with "transformer models" for years now told me "oh yeah, it's a bullshitter" when I tried my hand and got some truly bizarre, digressive, "addled", etc. output.

OTOH, it reminded me very much of my own mind (reinforced by ADHD, in my case).

This suggests to me, at least, that "the problem" isn't these models, per se. It's more like: these are probably only one module / layer in a system more similar to our brains. Just as scientists have identified distinct regions (more) involved in, say, language production, or (direct) visual perception, or etc., I'd suggest we've only just built the first substantially more practical / realistic hack / simulation (much like 3D game engines almost always use hacks - e.g., not even using the simple "Newtonian optics" model fully [i.e., "ray tracing"]) of a sort of language cortex. I'd further guess that it's going to take some maturation of a number of methods, technologies, etc. to realistically add more "cortices", but, I do think it's quite likely to happen in approx. the "decades" range...

Highly highly speculative - rather naively based on the way other technologies have developed and with a little basis in work I've done more directly in neurobio etc. No deep(er) reason / analysis, but, just my current very tentative hypothesis.

Not sure why you're being downvoted.

Are there other opinions about the cortex or module idea? Is there a fundamental problem with that idea I'm missing?

Hinton called them "confabulations" according to this:

https://www.technologyreview.com/2023/05/02/1072528/geoffrey...

I think "confabulation" is a way more accurate term, I wish it had stuck instead of "hallucination".

A hallucination is a problem with input. Confabulation is false output.

Confabulation is when a person mistakenly recalls details and tries to "fill in the blanks", without realizing what they are saying is untrue.

Maybe we could just call it babbling.