If you understand how LLMs work, that’s not surprising.

LLMs generate a sequence of words that makes sense in that context. It’s trained on trillions(?) of words from books, Wikipedia, etc. In most of the training material, when someone asks “what’s the name of the person who did X?” there’s an answer, and that answer isn’t “I have no fucking clue”.

Now, if it were trained on a whole new corpus of data that had “I have no fucking clue” a lot more often, it would see that as a reasonable thing to print sometimes so you’d get that answer a lot more often. However, it doesn’t actually understand anything. It just generates sequences of believable words. So, it wouldn’t generate “I have no fucking clue” when it doesn’t know, it would just generate it occasionally when it seemed like it was an appropriate time. So, you’d ask “Who was the first president of the USA?” and it would sometimes say “I have no fucking clue” because that’s sometimes what the training data says a response might look like when someone asks a question of that form.