story
"Hallucination" is a term that works well for actual intelligence - when you "know" something that isn't true, and has no path of reasoning, you might have hallucinated the base "knowledge".
But that doesn't really work for LLMs, because there's no knowledge at all. All they're doing is picking the next most likely token based on the probabilities. If you interrogate something that the training data covers thoroughly, you'll get something that is "correct", and that's to be expected because there's a lot of probabilities pointing to the "next token" being the right one... but as you get to the edge of the training data, the "next token" is less likely to be correct.
As a thought experiment, imagine that you're given a book with every possible or likely sequence of coloured circles, triangles, and squares. None of them have meaning to you, they're just colours and shapes that are in random seeming sequences, but there's a frequency to them. "Red circle, blue square, gren triangle" is a much more common sequence than "red circle, blue square, black triangle", so if someone hands you a piece of paper with "red circle, blue square", you can reasonably guess that what they want back is a green triangle.
Expand the model a bit more, and you notice that "rc bs gt" is pretty common, but if there's a yellow square a few symbols before with anything in between, then the triangle is usually black. Thus the response to the sequence "red circle, blue square" is usually "green triangle", but "black circle, yellow square, grey circle, red circle, blue square" is modified by the yellow square, and the response is "black triangle"... but you still don't know what any of these things _mean_.
When you get to a sequence that isn't covered directly by the training data, you just follow the process with the information that you _do_ have. You get "red triangle, blue square" and while you've not encountered that sequence before, "green" _usually_ comes after "red, blue", and "circle" is _usually_ grouped with "triangle, square", so a reasonable response is "green circle"... but we don't know, we're just guessing based on what we've seen.
That's the thing... the process is exactly the same whether the sequence has been seen before or not. You're not _hallucinating_ the green circle, you're just picking based on probabilities. LLMs are doing effectively this, but at massive scale with an unthinkably large dataset as training data. Because there's so much data of _humans talking to other humans_, ChatGPT has a lot of probabilities that make human-sounding responses...
It's not an easy concept to get across, but there's a fundamental difference between "knowing a thing and being able to discuss it" and "picking the next token based on the probabilities gleaned from inspecting terabytes of text, without understanding what any single token means"
