If an LLM were a person…

If a large language model were a person, they would be someone who has listened to a lifetime of conversations and learned how a good answer sounds. Ask them something and they answer readily. They are not opening a filing cabinet of checked facts. They are continuing a pattern that has usually been right.

That is also why they can be confidently wrong. A person who dislikes an awkward silence will sometimes fill it. A model does something close to that. When it does not have the fact, it still produces a sentence. We call that a hallucination. It is less a lie than a habit of keeping the conversation going.

Retrieval, the thing people shorten to RAG, is the practical correction. Before the person speaks, you hand them the folder: the document, the policy, the page that actually applies. They are still the same person. They are no longer answering from memory alone.

Agents, tools, and the newer ways of letting a model act are the same idea taken one step further. The person can now go and look, or ask a system, instead of only talking. Useful, as long as someone still decides what they are allowed to touch.

The question I keep is a simple one. Not whether the model is clever. Whether we gave it the right page, and whether we left it room to say that it does not know.