Why AI cannot hold a story structure

The model starts writing without knowing the ending

A language model does not plan and then write. It looks at the text so far, computes the probability of every possible next word, and samples one. Then it does it again, thousands of times. When it writes the first sentence, the ending does not exist yet.

Story structure is hostile to that. Building up means writing a scene that may be dull now so that a later scene can hurt. It sacrifices the present for a future point. Sequential prediction has no representation of that future point, so it takes the most natural next step every single time.

The result is not incoherence. It is flatness. Nothing breaks; the peaks simply disappear.

The probable continuation is always the moderate one

A real reversal means losing something irreversibly. That is not where the probability mass sits. After a conflict, the likely next words are reconciliation, explanation, understanding, compromise — because most stories that were ever written go there.

So the default arc becomes mild trouble, a conversation, a resolution. Perfectly readable. Nothing left behind afterwards.

Our own generation prompts contain repeated, explicit instructions about how far to build and how far to fall. Instructions you would not need if the model did it by itself. They are repeated because the model keeps drifting back.

It will quietly convert your ending into a kind one

We hit this in production. Conditions the author cared about most — no redemption, nobody ever finds out, it stays between these two — held at the outline stage and then reversed once prose generation started.

The cause is the same as above: stories that redeem their protagonist are the majority of the training distribution. The author instruction is a few hundred characters. The distribution is everything else. Left alone, the majority wins.

Our fix was to extract the inviolable intent of a work and re-inject it at every stage — outline, episode plan, prose. The assumption that saying something once is enough was simply wrong.

So can it still move a reader? Yes — and only for one reason

Everything above is a limit. It does not follow that AI-written fiction cannot move anyone.

What the model cannot do is decide how far to fall. That decision comes from a person: what would actually hurt to lose, who would be unbearable to be betrayed by. None of that is recoverable from a statistical average. It exists only in the life of whoever is writing.

Hand over those two lines and the enormous remainder becomes machine work: generating ten versions of a build-up scene, finding where to plant a detail, rewriting the same moment three ways. That remainder is exactly where most people used to give up. The constraint that kept stories unwritten was never the idea. It was the labour between the idea and a finished draft, and that constraint is gone.

Two lines are all you owe it

In practice you supply two things: what gets built up (whose trust, which place, what routine) and what is lost at the turn (stated as irreversible). The rest can be filled in.

Do not use adjectives. "Make it dramatic" retrieves the average of dramatic. Name the thing that is lost.