The use of “hallucinate” is a stroke of true evil genius in the AI world.
In ANY other context we’d just call them errors & the fail rate would be crystal clear.
Instead, “hallucinate” implies genuine sentience & the *absence* of real error.
Aw, this software isn’t shit! Boo’s just dreaming!
MacKay (2003, _Information Theory, Inference, and Learning Algorithms_, Cambridge University Press, <https://web.archive.org/web/20170610174915/https://www.inference.org.uk/itprnn/book.pdf>), calls them "spurious stable states": does that nicely avoid both sets of pitfalls?
> In ANY other context we’d just call them errors & the fail rate would be crystal clear.
The reason for using "error" and "fail rate" is that it doesn't anthropomorphize the AI.
@shanecelis @Catvalente
Confabulation has an existing technical meaning relating to creating false memories in response to brain damage or dementia. I don't think we want to use that to describe AI "hallucinations". I would call them "fabrications" or just "bullshit".
The AI slopware industry employs the best malign influence, ahem... marketing operatives in the world.
They're using all the standard ploys.
1. Fossil fuel fascist funding
https://www.bloomberg.com/news/articles/2024-11-06/saudis-plan-100-billion-ai-powerhouse-to-rival-uae-s-tech-hub
2. FOMO
https://www.bloomberg.com/news/videos/2026-04-05/why-you-should-ignore-ai-fomo-video
3. FUD
https://medium.com/@jung.marshall/fud-in-ai-ce3874cd99b1
https://om.co/2025/03/13/apple-intelligence-fud-dud-or-both/
4. Regulatory Capture
https://www.desmog.com/2025/12/11/the-koch-network-is-pushing-trump-to-accelerate-ai-documents-show/
5. Extortion & coercive user adoption tactics
https://www.forbes.com/sites/mattdurot/2025/07/17/bill-gates-charles-koch-and-three-other-billionaires-are-giving-1-billion-to-enhance-economic-mobility-in-the-us/
Saudi Arabia is planning a new artificial intelligence project with backing of as much as $100 billion as it seeks to develop a technological hub to rival the neighboring United Arab Emirates, people familiar with the matter said.
@Catvalente I keep telling people that they're just helping the scam when they use terms like "AI" to describe a thing that, by definition, can never ever become anything remotely close to AI. You're right. Using hallucination and any other terms the actual scammers come up with are all things that just help their scam instead.
This is why I almost always tell people not to call it "AI." Preferably refer to the actual individual tech involved specifically like "LLM" or "image diffusion." And yeah, don't refer to the inevitable errors as "hallucinations" because that term doesn't even make sense in context of how they function...
(To the people saying "technically it's not actually an error" it is a logic/correctness error, but, putting that aside, it sure AF is not a "hallucination")
@Catvalente Yes, I couldn’t agree more. In terms of mental phenomenology (if you’re going to use human terms to describe these things) the AI mistakes cannot be hallucinations. These mistakes are not based on a sensory experience.
A confabulation is a more accurate term - inventing things to fill the gaps
@DanDan420 @Catvalente In other words, AI hallucinations are not errors, that’s system working as expected.
The error is in between the chair and the keyboard.
@DanDan420 @Catvalente You're just describing the difference between a deterministic computer program and one with simulated probability (which are sometimes "models").
LLMs still run in the same way that other programs run on a computer...deterministically. However, the simulated randomness gives folks the impression that it is somehow different. It isn't different. If you play just about any computer game you'll have encountered what's going on here conceptually.
@DanDan420 @Catvalente Forget generated worlds, and think of the fundamentals: a pseudo random dice roll in an RPG. It's still deterministic ultimately (the randomness isn't really random), but it's close enough to real that it might as well be.
This randomness makes things wavy when you interact with the models because it's incorporated all over. That waviness (or temperature as they call it in some APIs) is what gives you unpredictable results (hallucinations) instead of consistent errors.
"However, the simulated randomness gives folks the impression that it is somehow different."
That is a really good point. I need to incorporate this when I talk with my students about LLMs.
@Catvalente
I hate using this kind of language when describing LLMs! I always want to post this diagram about de-anthropomorphising the output from chatbots (source = https://www.researchgate.net/publication/370842240_Mirages_On_Anthropomorphism_in_Dialogue_Systems)
The original response may seem okay at first but then you realize it's using first person language.
@callisto Nothing wrong with using the pronoun it but in the example I shared I think the word it might be confusing?
The phrase "This generative language model" might be a bit long, but replacing that whole phrase with just "It" seems confusing to me. What does it refer to?
But absolutely, using the pronoun it seems fine in many situtations.
@Catvalente From now on, my code has no bugs. It is designed to hallucinate, for which I should be made a billionaire.
Sure. I could get behind this.
@Catvalente yes! So much BS.
You may appreciate this: https://berryvilleiml.com/blog/
See esp the Anthropic entries
@Catvalente @jwz see also.
TL;DR given the way all outputs are generated, if one is a hallucination, they all are.
I’m sure many other have made this observation, but even just reading this post without reading the linked article made me realise (or remember) that... *All* LLM output is, in fact, a hallucination. Because the way it formulates a “hallucination” *is exactly the same* as how it formulates a response *we don’t consider* a hallucination. Same with “good” vs “bad” summaries (and whatever the relative occurrence of each is). #NoAI #HumanMade
Even the term 'lie' or 'prevaricate' (IT IS A PIE) is so fundamentally inaccurate to what the program does, though we've used that anthromorphization in other, non-chatbot ways through the years.
Of course, the robot does not lie: it outputs what it has been told to output. The programmers and executives, on the other hand...
I call them lies.
I wanted to check on a detail about an event in a book because I want sure it was appropriate for a minor. Instead of rereading the book, I foolishly asked Chat GPT. The LLM said that event didn't occur.
I knew it did, I just couldn't remember a specific detail. I kept pressing and eventually it responded that it knew about the event but said it didn't occur because it didn't want to upset me. (Because gifting a minor a book with traumatizing content wouldn't be upsetting!?)
So, yeah, that a lie in my book.
@Catvalente The other thing with this is that "hallucinate" implies that producing accurate output without fabrications is the normal way an LLM works, and "hallucination" is some failure of this that may be treatable.
No. Fabricating linguistically plausible output is exactly what we expect a "language model" to do. The technology, with some tuning, may do surprisingly well at producing accurate output, but it still is fundamentally a model of language, and there's no particular reason to believe it's possible to make a version that isn't prone to hallucination without a completely different methodology than how LLMs work.
I've simply called it "lies" forever.
The argument always is the same - it can't be lies, it's just wrong.
Well then, it can't be a hallucination. If we're going to pretend it's intelligent, and it tells me things that aren't true *with confidence and assurance* - that's a lie.
(In reality, now, when using it for code assist, I just call it stupid. Nobody argues about that.)
@Catvalente They play these games all of the time with language and it's horrible to watch everyone get sucked in and start using the lingo without knowing the obvious design behind it.
Another example of this is "hyperscaler".
And perhaps more relatedly "agent"...which as Ed Zitron points out...means a chatbot.
Yep, its probabilistic computing, using statistical methods over billions of parameters.
@Catvalente The so called hallucinations are in fact called beam search.
Beam search is like trying say the top 10 most -likely- words, and doing that several steps deep to construct the most likely phrase.
@painting_squirrel @Catvalente yes. this.
But worse: I wouldn't say "sometimes by random chance." The results do _frequently_ match reality because the statistics driving the output generation were based on text that matched reality. But when they don't match they still sound indistinguishable without checking against objective reality.
So it is a lull-you-and-kill-you game.
@poleguy @Catvalente
I'll concede your point but only because "sometimes" and "random chance" are very badly defined terms that are not very intuitively understood by humans (myself included).
To wit: only because something is statistically well tuned to give good results most of the time does not mean that the results are not inherently random. "random" != "everything has the same probability"
The best term, IMHO, is baloney (or BS).
It's not a lie because the AI cannot have intent to deceive.
It's not a hallucination because the AI cannot be deceived.
It's not an error or bug or mistake because it's doing what the creator intended.
When you input a question to an AI, the output of an AI is baloney. And that's OK if you're willing to eat baloney.
Stephen I Moore
Catherynne M. Valente
Giflian
Can Acar
Chris Goss
Marko Lindqvist
Dan Hatton
Bloodymirova
I am Jack's Found 404
toolbear#🌶️
🔥 Mireille Sillander 🔥
James Baillie
Max
👻™ supreme goblin 🐊
Shane Celis
Roger Moore
Nicole Parsons
J. Steven York
Nazo
drdr6262
bituur esztreym
DanDan420
slotos
Avocado Toast
Pascal Zaugg
Ahimsa
Callisto
Steve's Place
noplasticshower
thedoh 🇨🇦
europlus 
Dominik Grabiec
Jennifer Kayla | Theogrin 🦊
Jennifer
Ian Douglas Scott
Assimilateborg
Tom Bortels
Tim
Yet another Josh 
Matthew Merkovich
Dodo
spdrnl
DJGummikuh
Multi Purr Puss 
Nini
Stoneface Vimes
😷nio
Orava
poleguy looking for lost tools
無
killick
We don't need AI. AI needs us.