Right, they just make all things up by design, no matter what you ask about. There is a statistical chance that a pattern of words output may correspond to fact but just as a good a chance not. The LLM's literally know nothing about the world, its just statistical word pattern output.

Lies and falsehoods are just as valid sentences as the truth.

I got so many links to github repos which didn't exist lol

Somebody went as far as contacting the author of a paper based on chatGPT suggestion. Of course the author was pretty sure he never heard of that paper.

"They do give sources if you ask for them... but have a habit of inventing plausible sounding but completely fake sources."

Basically, the way you want to think about it is, no, they can not give sources. That information is not in their neural net. It can't be. There isn't anywhere to encode or represent it.

What they can do is make a guess what a source might look like, but even if they are right, it is only because they happened to guess correctly, not because they knew the source. They don't. They can't.

It isn't that "they give sources but they might be wrong", it is "they will make up plausible-sounding sources if you ask just as they'll make up plausible-sounding anything else you ask for, and there's a small chance they'll be right by luck". For more normal factual-type questions part of the reason they are useful is that there's a good chance they'll be right by what is still essentially luck, but for sources in particular there's a particularly small chance, by the nature of the thing.

That said with enough training and training data, the line between "plausible sounding" and "accurate" gets thinner and thinner. This will be especially true as these AI models refine their results based on user interactions. Being right for the wrong reasons becomes less and less relevant the higher the accuracy goes up, and at a certain point, it might get so good that no one cares.

Maybe human intelligence is more like that than we're willing to admit ;)

Source identification isn't a space amenable to guessing, no matter how much data you throw at it.

Here's an exercise you can try: Cite some information from the next issue of Science to be published. Cite anything you like from it.

You can make some plausible stuff up. You could make even more plausible stuff up if you went and scanned over the past few issues first. But without specific knowledge of the contents of the next issue, you aren't going to be able to create real citations. This is what LLMs lack, by their nature. It's not a criticism, it's a description.

You can't guess sources. The possibility space is too large, the distribution too pathological, and the criteria for being correct too precise.

GPT will never cite sources correctly. Some future AI that uses GPT as a component, but isn't entirely made out of a language model, will be able to, by pulling it out of the non-GPT component. Maybe it'll need to be built as an explicit feature, maybe it won't, only time can tell. But expecting language models to cite sources correctly is not sensible. It's just not a thing they can do.