The difference is knowledge. You know what an apple is. A LLM does not. It has training data that has the word apple is associated with the words red, green, pie, and doctor.
The model then uses a random number generator to mix those words up a bit, and see if the result looks a bit like the training data, and if it does, the model spits out a sequence of words that may or may not be a sentence, depending on the size and quality of the training data.
At no point is any actual meaning associated with any of the words. The model is just trying to fit different shaped blocks through different shaped holes, and sometimes everything goes through the square hole, and you get hallucinations.