Why every AI character sounds the same

There is no memory of the character

The sense that a character settles in as a conversation goes on is an illusion. Nothing persists on the model side; each turn re-reads the description and the history together. The tenth reply is produced by exactly the same procedure as the first.

So when a character drifts, it is not because something was forgotten. It is because the input for that turn differed from the input before. Consistency is a property of your input, not of the model.

Getting this backwards leads to the wrong fix — adding more description — which usually makes it worse.

Character description is weaker than you think

Some real numbers. The instruction block we send to generate one episode runs to tens of thousands of characters: structure, scene content, prohibitions, style. Of that, 300 to 450 characters concern how a character speaks and behaves. Well under one percent.

Worse: the variation between two runs with identical settings can exceed the variation between two different settings. When a change seems not to take effect, often it did take effect and was buried under sampling noise.

This does not mean description is useless. It means adding volume to a weak signal does not strengthen it. You have to change the kind of information, not the amount.

Adjectives retrieve the average of that adjective

Write "cold and aloof" and you get the centroid of cold and aloof: clipped answers, averted eyes, short sentences. That is not your character. That is the centre of mass of every aloof character in the training data — which is precisely why they all converge.

An adjective is an instruction to be average. Replace it with a choice: what this person does in a specific situation. Choices have no centroid, so they move the distribution.

Prohibitions make dialogue wooden

Another production failure. When dialogue started reading like a translation — narration fine, speech unnatural — our first move was to add banned phrasings. It got steadily worse.

Under heavy constraint the model retreats to literal compliance: satisfying the rules takes priority over sounding human. You end up with lines that violate nothing and that nobody would say.

What fixed it was subtraction: remove the bans, keep one positive line — speak with the same naturalness as the narration. Prohibition-shaped instructions almost always overcorrect.

A character stands up when one line comes from a life

None of this means AI characters cannot be memorable. The condition is narrow and specific: the writer knows something concrete about this person.

"Leaves the room when thanked" does not come out of an average. Only someone who watched a particular person, or is that person, writes it. And one such line is enough — the thousand scenes around it can be generated.

Writing distinct characters used to be a craft skill: years of drafting, cutting, revising. With that work externalised, what is left is whether you have looked closely at anybody. Plenty of people who never learned to write have been watching carefully their whole lives. Until now that never became a book.