Mental Model: Author, Not Character

Knowledge

You have now seen how an LLM works: tokens, embeddings, attention, training. That is the mechanics. But next time you chat with ChatGPT or Claude, you need a picture in your head — a mental model that helps you make sense of what you have learned.

The most useful picture is that of an author writing a character.

When we chat with an LLM, the response sounds like it comes from a person. Our brain almost automatically turns it into a counterpart — "the assistant," "Claude," "ChatGPT." That is a convenient simplification, but it is misleading. Because that counterpart does not exist. What exists is an author writing a character in real time.

The metaphor: Imagine a novelist bringing a character to life. With each sentence, the author asks: "What would this character say next?" The author has no knowledge of the character's world — only a feel for what fits, based on everything they have ever read. The LLM is the author. The helpful assistant persona you are talking to is the character.

So when the model writes "The capital of France is Paris," it did not "look it up." It reproduced a pattern that was extremely stable in the training data. The pattern matches reality — very reliably, in fact. But the process is the same as for any other answer: the author writes what fits the character.

Understanding

Every answer comes about the same way

That is the core idea. Whether an answer is factually correct or not, the mechanism behind it is identical. The model does not verify truth, it picks tokens. The picks come out the way the training data suggests.

This picture quietly explains a lot of things that would otherwise seem puzzling:

• Why do role-play prompts work? ("You are an expert in Roman history...") Because the author can play any character. You tell them which role is on now, and they write accordingly.
• Why do wrong answers often sound more convincing than right ones? Because the author writes both equally well. Their job is not to be right — their job is to make the character sound coherent.
• Why does the answer vary when you ask the same question twice? Because the author writes fresh every time. There is no stored "answer" — only the process that produces answers.

What this means for you

The model is not unreliable. It is reliable in a different way than you might first think. An LLM answer is not a dictionary entry that someone looked up. It is a draft — like the first draft from a helpful colleague who has read a lot and writes quickly.

You do not read a draft with suspicion. You read it with interest — and with your brain engaged. You take what is useful and notice when something does not quite fit. That is exactly the right stance for LLMs.

Apply

The mental model leads to a stance that is surprisingly relaxed: fundamental questioning. Not suspicion, not defensiveness — the same attentive reading you give to a colleague's draft or a Wikipedia article.

Three small questions that help:

1. Could this also be different? Where is the answer unambiguous, and where could the author have written something else? For "Paris is the capital of France," there is zero room. For "The three most important reasons for X's success are...," there is a lot.
2. Where might this come from? Does the answer sound like common knowledge, a specific source, or a quote that no one would recognize? Specific numbers, legal sections, and names are the most common stumbling blocks.
3. What would the test be? How would you find out if it is true — without asking the same model again? Sometimes a minute on Google is enough. Sometimes you already know best yourself.

Reflect

If you keep the picture of the author in your head, it does not change what you can do with LLMs — but how you read the answers. That is a small difference with a big effect. The rest of this module — especially the next section on the limits of LLMs — builds directly on this foundation.