How to Use AI for Creative Writing

In 2024, Anil Doshi and Oliver Hauser ran an experiment that produces the most useful single fact about AI and fiction. Published in Science Advances, it gave some writers story ideas from a language model and others nothing, then had the resulting short stories evaluated.

The AI-assisted stories were rated more creative, better written and more enjoyable — and the effect was largest for the writers who scored lowest on creativity to begin with.

Then the second finding: the AI-assisted stories were measurably more similar to each other than the stories written without it.

The authors describe it as a social dilemma. Each writer is individually better off. Collectively, the range of what gets written narrows.

For you, sitting alone with a draft, that translates into something concrete. The model will raise your floor and lower your ceiling at the same time. It reliably makes bad prose competent. It also pulls competent prose toward the middle of everything it has read — and the middle is precisely where nothing memorable lives.

Which suggests an obvious division of labour.

Use it on everything except the prose

The tasks where AI helps a writer are the ones where convergence toward a well-formed norm is a feature:

• Structural diagnosis. Where does the tension drop in this chapter? A model is a decent proxy for an inattentive reader, and inattentive readers are the ones you lose.
• Continuity. Tracking whether the character with grey eyes in chapter two still has them in chapter nine. Mechanical, tedious, and genuinely error-prone for humans.
• Options generation. Fifteen loglines, ten titles, six ways the scene could go next. You are not looking for the model's best idea. You are using volume to find the one that makes your idea move — and then discarding the other fourteen.
• The query letter, blurb and synopsis. These are commercial forms with conventions. Converging on the norm is the entire job.

And the task where it hurts:

• Generating the sentences a reader will actually read.

Not because AI prose is bad. Because it is average — competent, rhythmically uniform, reaching for the phrase that most often follows. Voice is deviation from the expected, and a model trained to predict the expected cannot manufacture it. It can only sand it off.

That split is how the creative writing pack is built: brainstorming, outlining, character and world consistency, developmental critique, line-edit diagnosis — with the prose prompts pointed at your text, converting your telling into showing or fixing your flat dialogue, rather than producing new pages from nothing.

Ask for diagnosis, refuse the rewrite

The most valuable prompt in creative writing is one that forbids the model to write.

Do not rewrite any of this. Tell me: where did your attention drift, which paragraph is doing the least work, what am I explaining that I should be dramatising, and what do I seem to be avoiding? Point at line numbers.

The moment a model offers a rewrite, two things happen. You lose the diagnosis — you now have a fix without knowing what was broken. And you start editing toward its solution instead of yours, which is exactly the homogenising force the Doshi and Hauser result identifies.

Get the diagnosis. Close the window. Fix it in your own words.

Anchor it to your own voice, not good writing

When you do want the model touching prose — a line edit, a POV change, a tense conversion — the instruction that matters is not make this better. It is a fixed point:

Preserve my sentence rhythms and my word choices. You may cut, reorder and tighten. You may not replace my vocabulary with more common alternatives or smooth my sentence lengths toward uniform. If a line is odd on purpose, leave it.

Left unconstrained, a model's default edit is regression to the mean: it evens out sentence lengths, replaces the unusual word with the expected one, and resolves deliberate roughness. Every one of those is locally an improvement and cumulatively the death of a voice.

A useful check: keep the original, and after any AI-assisted edit, read both aloud. If the edited version is easier to read and harder to remember, you have traded the wrong way.

The idea you didn't have

The sharpest use of the Doshi and Hauser finding is defensive. If a model can produce your premise in five seconds, so can every other writer with the same tool — and the study says their stories will resemble yours.

So invert the ideation prompt:

Give me the twenty most likely directions a story with this premise would go. I want to know the obvious ones so I can avoid them.

That extracts the value the model is actually good at — knowing what is common — without adopting it. The convergence becomes a map of what to avoid rather than a set of suggestions to accept.

The part that stays yours

There is a version of this argument that says use AI for none of it. That is not what the evidence supports. The individual gains in the study were real, and the writers who gained most were the ones with least craft — the people for whom the alternative was not a more original story but no finished story.

The honest position is narrower. Use it for structure, for consistency, for volume, for the commercial packaging, and for honest diagnosis of what isn't working. Then write the sentences yourself, because those are the only part a reader will remember, and the thing that makes them memorable is the thing a model is built to remove.

---

Source: Doshi, A. R. and Hauser, O. P., Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances 10, eadn5290 (12 July 2024).