Internally, LLMs know a whole lot more about the truth and uncertainty of their prediction than the say. Pushing that to words is difficult but not impossible.

https://news.ycombinator.com/item?id=41504226

Some of the text that the LLM is trained on is fictional, some of the text that its trained on is factual. Telling it to not make things up can tell it to generate text that’s more like the factual text. Not saying it does work, but this is a reason how it might work.

The model can be trained to interpret "don't hallucinate" as "refer only to the provided context and known facts, do not guess or extrapolate new information", which wouldn't get rid of the issue completely, but likely would improve the quality if that's what you're after and if there's enough training data for "I don't know" responses.

(But it all depends on the fine-tuning they did, so who knows, maybe it's just an Easter egg)

I think it's more likely that it's included for liability reasons.

I did something similar and to my surprise effectively made the LLM in my tests admit when they don't know something. Not always but worked sometimes. I don't prompt "don't hallucinate" but "admit when you don't know something". It's a logical thing in the other hand, many prompts just transmit the idea of being "helpful" or "powerful" to the LLMs without any counterweight idea. So the LLM tries to say something "helpful" in any case.

Playing around with local models, Gemma for example will usually comply when I tell it "Say you don't know if you don't know the answer". Others, like Phi-3, completely ignores that instruction and confabulates away.

That's not a big problem since it will understand it, and if they already fine tuned the model to work with that prompt it'd get harder to change.