That seems overly emotional, to lie requires intent right? It implies the will to deceive. We don't accuse people of deception when they are simply wrong after all, so it feels inappropriate to accuse a machine that is predicting the next most likely word of lieing.

There is actually some basis in concept. Emergent deception ...

https://bounded-regret.ghost.io/emergent-deception-optimizat...

Much obliged! That article is off with a strong start

Yes. Lying requires being able to discern truth from falsity. But LLMs don't know what's true even when they say true things. That's why "hallucinate" is a better word ... better but not perfect because for an LLM it's all hallucinations all the time. Some of those hallucinations (most of them) happen to turn out to be true.

Lying requires intent and knowledge.

And hallucination requires consciousness and sentience. There really needs to be better terminology for this.

Consciousness? I dunno about that. Does an ant on the right drugs not hallucinate?

I agree though that it's a bad term for what current LLMs do. "Confabulate" seems closer.

> Does an ant on the right drugs not hallucinate?

I think we have no way of knowing, just like we don't know if an ant has consciousness or not.

There's neuroscience research on hallucinations, for example https://www.math.utah.edu/~bresslof/publications/01-3.pdf "What Geometric Visual Hallucinations Tell Us about the Visual Cortex". The word "conscious" does not appear.

> The results are sensitive to the detailed specification of the lateral connectivity and suggest that the cortical mechanisms that generate geometric visual hallucinations are closely related to those used to process edges, contours, surfaces, and textures.

Now this example is only one theory, but it shows that serious people can talk in these terms about hallucination and expect to be understood and not have to argue past reviewer #2, without this whole question vitiating their discussion.

Also, btw, we don't fully know that GPT-4 is not conscious, if you really think it matters. When the question comes up, people who are certain it's not usually point to the fact that it's not running continuously, as if that had any bearing on the question.

Statistical fallout of unexpected words droppings.

We humans call it "misspeaking".

Speak quickly, often enough, and without sufficient training and yule misspeak.

"Wrong", "a fail-state", "I'm sorry, Dave". Any of these do anything for ya?

Non consistent databases is how i’d call these chat toys. We made software to get accurate results but now they want to sell us the idea that a system that consistently produces inaccurate results is good … because it “hallucinates” just like hoomans. As if that’s what we needed from machines in the first place, to mimics our limitations.

confabulate: to replace the gaps left by a disorder of the memory with imaginary remembered experiences consistently believed to be true (Collins)

They neither lie nor hallucinate. They get it wrong. It's a fail.

Is the guy who adds an edit whilst citing a seemingly credible source also a liar? No, they're wrong.

Intent is everything. AI does not have intent. Don't feed the animals, don't personify the LLM.

Well, in some cases, negligence is sufficient for a lie or deception. E.g. forgetting to properly cite results is plagiarism, whether you intended to lie or not. Maybe it makes sense to hold the owner of the hardware that produces these <hallucinations/lies/failures/deceptions/confabulations> accountable for it. In many places with freedom of speech, this will be difficult probably.

Words can take new meanings when they are used in other, sometimes technical, contexts.

I think this is more accurate for non-emotional reasons. The GPT models have actually been trained to lie in the sense that they will tell you something convincing not just about the topic, but also how they are getting the information, e.g. “I checked my sources and found…”. You can also see jailbreaks where it has been trained to reverse belief on sensitive topics. This is consistently deceptive bias rather than random variance.

Imagine if Apple had a generative image model that produces fake metadata and watermarks showing the data cam from a real iPhone, and then signs it with their key. That would show obvious intent to deceive rather than attempt to accurately represent reality.

I think they bullshit is a better way to put it.