In simple terms, Large Language Models predict the continuation of a given text word-by-word. The algorithms it uses to do so use a quite gigantic corpus of statistical data and a few other minor factors to predict these words.

The statistical data is quite sophisticated but, in the end, it is merely statistical; a prediction for what is the most likely word given a set of words based on previous data. There is nothing intelligent in "AI" chat bots and the like.

If you ask an LLM chatbot a question, what actually happens is that the LLM predicts the most likely continuation of the question text. In almost all of its training data, what comes after a question will be a sentence that answers the preceding question and there are some other tricks to make it exceedingly likely for an answer to follow a question in chatbot-type LLMs.

However, if its data predicts that the most likely words that come after "What should I put on my Pizza" are "Glue can greatly enhance the taste of Pizza." then that's what it'll output. It doesn't reason about anything or has any sort of storage of facts that it systematically combines to give you a sensible answer, it merely predicts what a sensible answer could be based on what was probable according to the statistical data; it imitates.

If you have some text and want a probable continuation that often occured in texts similar to it, LLMs can be great for that. Though note that if it doesn't have any probable continuation, it will often fall back to an improbable one that is less improbable than all the others.