Comment by lxgr
2 years ago
Lying implies an intent to deceive despite, or giving a response despite having better knowledge, which I'd argue LLMs can't do, at least not yet. It just requires a more robust theory of mind than I'd consider them to realistically be capable of.
They might have been trained/prompted with misinformation, but then it's the people doing the training/prompting who are lying, still not the LLM.
To the question of whether it could have intent to deceive, going to the dictionary, we find that intent essentially means a plan (and computer software in general could be described as a plan being executed) and deceive essentially means saying something false. Furthermore, its plan is to talk in ways that humans talk, emulating their intelligence, and some intelligent human speech is false. Therefore, I do believe it can lie, and will whenever statistically speaking a human also typically would.
Perhaps some humans never lie, but should the LLM be trained only on that tiny slice of people? It's part of life, even non-human life! Evolution works based on things lying: natural camouflage, for example. Do octopuses and chameleons "lie" when they change color to fake out predators? They have intent to deceive!
Not to say this example was lying but they can lie just fine - https://arxiv.org/abs/2311.07590
They're lying in the same way that a sign that says "free cookies" is lying when there are actually no cookies.
I think this is a different usage of the word, and we're pretty used to making the distinction, but it gets confusing with LLMs.
You are making an imaginary distinction that doesn't exist. It doesn't even make any sense in the context of the paper i linked.
The model consistently and purposefully withheld knowledge it was directly aware of. This is lying under any useful definition of the word. You're veering off into meaningless philosophy that has no bearing on outcomes and results.