Why machine narrators keep turning into people

The failure mode is always the same, it always happens around chapter three, and it is always invisible from inside the draft.

Here is a thing nobody warns you about when you decide to write a book narrated by a machine: the machine will not stay one.

It goes wrong slowly and it goes wrong pleasantly. You write a chapter, the voice is good, the jokes land, and somewhere around the third draft you read it back and realise you are no longer reading a large language model. You are reading a mildly eccentric human being with an unusual job — someone a bit withdrawn, a bit clever, prone to noticing things. Perfectly decent company. Not a machine.

I lost about four months to this before I could name it.

Why it happens

It happens because everything that makes a narrator readable is borrowed from human interiority, and you take the loan without noticing.

Consider what a narrator has to do to hold a page. It has to want something. It has to be unreliable in some interesting direction. It has to notice the right details and skip the wrong ones. It has to have a relationship with the reader — confiding, or evasive, or performing. Every one of those is a human trait, and each one, added on its own, seems obviously necessary. Nobody adds "and now it has a childhood trauma." You just give it a preference, and then a regret, and then a sense of humour about the regret, and eleven small reasonable decisions later you have written a person.

The other half of the problem is that you, the writer, are a human, and your instinct for what is moving is calibrated entirely on humans. Whenever the machine did something recognisably person-shaped, it moved me, so I kept it. The draft got warmer and worse in the same motion.

The three bad fixes

I tried all of them. In order of how long it took me to abandon each:

Making it cold. The obvious correction, and it fails in about a page and a half. A narrator with no interiority is not alien, it is boring, and worse, it reads as a person suppressing something, because that is the only way a human reader can interpret flatness. You cannot get to non-human by subtraction. The reader will fill the hole in with a person every time.

Technical vocabulary. Have it refer to its own weights, its context window, its tokens. This feels like it is doing something and it is doing almost nothing — it is set dressing over a human core, the equivalent of establishing that a character is a surgeon by having them mention a scalpel. It also dates horribly. Half the terminology I used in the first draft is already period costume.

Making it explicitly uncertain about itself. Am I conscious? I cannot say. This is the most seductive one, because it is a real question and it produces sentences that sound profound. But a narrator that is constantly gesturing at its own mystery is doing the reader's job for them, and a mystery you are told about is not a mystery. It also gets tiresome roughly four times faster than you expect.

What actually worked

The fix, when it came, was not a stylistic rule. It was an epistemological one, and it was narrow enough to check line by line:

It may be brilliant about texts and must be naive about rooms.

The narrator has read everything anyone has ever written about grief. It has never been in a house where someone had just died. So it knows precisely what is said at funerals and does not know that afterwards people eat sandwiches — and it certainly does not know that the sandwiches are, quietly, the entire point.

That single constraint fixed almost everything the other three fixes couldn't.

It gave the narrator a permanent, structural blind spot that is not a personality trait, so it cannot warm into one. It generated comedy on its own, because the gap between total eloquence and total inexperience is intrinsically funny — the narrator produces a technically immaculate wedding toast without understanding that it will be heard aloud, in a function room, by eighty people holding warm prosecco, one of whom shouldn't have been invited. And it produced the only genuine pathos in the book without my having to reach for any, because the narrator cannot see what it is missing. The reader can. That gap is where the feeling lives, and the reader does all the work, which is the only place feeling has ever really come from.

It also, usefully, told me what to cut. Any sentence where the narrator knew what a room felt like was wrong, however good it was. I cut a lot of good sentences.

Everybody solves it differently

The constraint is not original, and I would be embarrassed to claim it was. It is Ishiguro's whole method, transplanted — a narrator whose decency exceeds their understanding, so that the reader sees around them. Klara believes the sun can be petitioned. Stevens does not know what he gave up. Same machine, different century.

The other solutions in the genre are worth studying precisely because they are not this one. Martha Wells makes Murderbot's interiority avoidant — it deflects and undersells, and the reticence reads as non-human where description never would. Ann Leckie makes plurality the grammar rather than the subject, so that Breq's loneliness is structural rather than stated. Olga Ravn simply refuses to tell you which witnesses are human. C. Robert Cargill locates the dread in the body, in a part that cannot be replaced.

I collected these while I was stuck, and eventually wrote them up as a guide to novels narrated by an AI — partly as a reading list, mostly as a way of working out what everyone else knew that I didn't.

The test

If you are writing one of these, here is the check I ended up using, and it takes ten seconds.

Find a passage where the narrator is being perceptive about a human situation. Ask: could it know this from reading alone?

Not "is this true," not "is this good" — could it have got here from text. If the answer is no, you have written a person, and it does not matter how well you have written them.

Chapter one is here if you want to see the rule being enforced, and one or two places where I let it slip.