Theory of mind: why the best stories are about who knows what

What theory of mind is

Psychologists test it with a marble. Sally puts the marble in her basket and leaves the room. Anne moves it to her box. Sally comes back. Where will she look?

To answer the basket, you have to hold two versions of the world at once: where the marble is, and where Sally thinks it is. Most children manage it from about the age of four. Younger children tend to answer with where the marble really is, because for them there is only one world, the true one.

Fiction is a machine for knowledge gaps

Stories are built out of that double bookkeeping. The literary scholar Lisa Zunshine argues that we enjoy fiction because it keeps inviting us to guess at other minds, which our brains do all the time anyway (Why We Read Fiction, 2006). Look at what the oldest devices of storytelling are made of:

• Dramatic irony. The audience knows something a character does not. In Romeo and Juliet we know Juliet is only asleep; Romeo does not, and acts on what he believes.
• Suspense. The audience knows the danger is coming, and the characters do not.
• Mystery. Nobody knows yet, and the reader finds out alongside the detective.
• The lie and the betrayal. Someone acts on a false belief that someone else planted, or on a loyalty that was never real.
• Minds inside minds. She suspects that he knows that she lied. Readers follow several of these layers without noticing they are doing it.

Some researchers went further and claimed that reading literary fiction trains this ability. A 2013 study in Science reported exactly that. When the same authors ran three preregistered replications, the results were mixed: one succeeded, and two failed without settling the question (Kidd and Castano, 2019). In two of those experiments, readers also judged characters in popular fiction more predictable and more stereotyped than characters in literary fiction. Whether or not fiction trains theory of mind, it runs on it.

Why an AI narrator gets it wrong by default

A language model writing a story sees everything in its context at once: the villain's plan, the heroine's secret, the name of the stranger you have not met yet. Nothing in that text says who was in the room when something was said. So its default is an omniscient cast, where every character knows what the prompt knows.

In our own playtests this month we met all three classic leaks: a name the hero had never heard, a hidden truth written plainly into the prose, and a twist given away before its scene.

This is not because models cannot reason about beliefs. Given a classic false-belief puzzle, recent models often solve it: a 2024 study found GPT-4 solving 75% of its tasks, about as well as six-year-old children (Kosinski, PNAS). Others showed that small changes to the same puzzles can make models fail (Ullman, 2023), and the debate is still open. In a story, the problem comes before any reasoning. The narrator is handed everything, and nothing separates what each character has actually witnessed.

What it opens up for a player

When a story keeps track of who knows what, a kind of play appears that an omniscient cast makes impossible:

• The bluff. You lie to a guard, and the lie holds, because he has no way to know better.
• The secret. What you hide stays hidden until you, or the story, reveal it.
• The investigation. Clues are found, not recited. Nobody hands you the answer because the narrator happens to know it.
• The misunderstanding. Someone acts on a false belief, and the consequences are real.
• The betrayal. An ally knew all along, and there is a scene where you realise it.
• The surprise. The story knows something you don't, and keeps it.

What we are building, in broad strokes

Nostyss keeps track of who has met whom and who has learned what, and writes each scene from inside what your character knows. A secret stays out of the prose until the story reveals it. Someone you have not met is described rather than named, until the story introduces them.

This is recent work, and it is not finished. It matters to us more than any single feature, because it is the difference between a story that remembers and a story with people in it.

What still breaks

• Inference. A sharp character should be able to guess. We would rather she guessed than simply knew, and drawing that line is hard.
• News off the page. Rumours travel between scenes, and a story has to decide what reached whom.
• The second order. What she believes you know is harder to keep straight than what she knows.
• The suggestions. The choices we propose can still name someone you have not met yet.

None of these is fixed by a bigger model. They are fixed, when they are, by the story keeping better books.

Why this is the fun part

A story is not only what happens. It is who knows it, when, and what they do about it. Give the player a mind among other minds, and the oldest pleasures of fiction come back: the held breath, the lie that works, the reveal that lands.

To check this on any AI story app, the fourth of our five memory tests does exactly that: does a character know something they never learned?

See whether it actually remembers

Ten worlds you can start without an account. Play a few chapters, contradict something you established earlier, and watch what the world does about it. That test is the honest way to judge any of this.

Keep reading