Guide

Does AI make writing better? What the studies actually found

Faster and, on average, better. Also more alike. Two peer-reviewed experiments explain both halves, and together they say something specific about how to use a model on a book.

Neubook Write · 5 September 2026

Study one: faster and better, on average

Shakked Noy and Whitney Zhang at MIT gave 453 college-educated professionals real writing tasks from their jobs (press releases, reports, analyses) and let half use ChatGPT. Their paper in Science, Experimental evidence on the productivity effects of generative artificial intelligence, found the treated group finished 40% faster with output graded 18% higher by blind evaluators. The most interesting detail is who gained: writers who scored lowest in the first round improved the most, so the gap between weaker and stronger writers narrowed.

Two caveats matter for authors. The tasks were short, 20 to 30 minutes each, and the participants mostly used the model to produce a draft they then polished. Nobody in the study wrote a 40,000-word book and hoped the model's voice would hold for 12 chapters.

Study two: more creative, and more alike

Anil Doshi and Oliver Hauser asked around 300 people to write short stories; some got no help, some got one story idea from a model, some got five. In Science Advances, Generative AI enhances individual creativity but reduces the collective diversity of novel content, they report that AI-assisted stories were rated more creative, better written and more enjoyable, with the biggest gains for the least creative writers. But the AI-assisted stories were measurably more similar to one another. Everyone's helper suggested the same few directions, so the pool of stories shrank toward a centre.

What both studies say when you put them together

A model raises the floor and lowers the ceiling of variety. Left to its own devices it will produce a competent book that resembles every other competent book on the topic. That is fine for a compliance manual and fatal for a memoir, a devotional or anything a reader buys for the author's company.

Follow-up work has been testing remedies. A 2026 paper in Computers in Human Behavior: Artificial Humans found that giving the model a distinct persona reduces the homogenisation effect, and an ex-ante evaluation of idea-diversity collapse argues for measuring diversity before generation rather than after.

Five rules that follow from the evidence

Supply the idea; let the model supply the hours. Noy and Zhang's gains came from drafting and polishing, not from ideation. Doshi and Hauser's homogenisation came from letting the model choose the idea.
Feed it your specifics before it writes a word. The interview matters more than the prompt. Who is this for, what happened to you, what do you disagree with.
Give it your voice, not a "professional" one. A sample of your own prose acts as the persona that the 2026 study found protects variety.
Edit for difference, not just correctness. Ask of every chapter: could another author on the same topic have written this paragraph? If yes, add the thing only you know.
Use the time you saved on the parts a model cannot do. Talk to readers, check facts, and write the one chapter that is actually yours from scratch.

This is how Neubook Write is built. Two interview questions and an optional voice sample come before the plan; the model drafts and edits, and you approve every chapter. Your first book is free to read and download.

Start a book, free

Does AI make writing better? What the studies actually found

Study one: faster and better, on average

Study two: more creative, and more alike

What both studies say when you put them together

Five rules that follow from the evidence

Sources

Related