If You Use AI, Will Your Novel Still Sound Like You?

What Novelists Actually Told Cambridge

The best evidence on how novelists feel comes from a report by Clementine Collett at Cambridge’s Minderoo Centre for Technology and Democracy, published in November 2025. She surveyed 258 published UK novelists between February and May that year. The full report is worth an hour.

Asked how often they use generative AI in their work, 67% said never, 22% rarely, 10% sometimes and 1% often. So when you read that a third of novelists use AI, most of that third means rarely. The commonest uses were looking things up (20%), editing text they’d written themselves (8%), and marketing copy (5%).

The share who felt “extremely negative” about each use falls away as the task moves from the sentence to the chore:

Use of AI	Novelists “extremely negative”
Writing a whole novel	97%
Writing longer sections	92%
Writing short sections	87%
Editing text written without AI	43%
Sourcing historical facts	33%
Sourcing general facts	30%

There’s more than taste behind this: 59% said they know their work has been used to train AI models, and nearly all of those said they gave no permission and got no money. And 85% expect their future income to fall.

Treat the survey as a good picture of one group. It’s British, 90% of those novelists are traditionally published, and 78% of all respondents are women. A romance writer self-publishing three books a year may see it differently, and the report includes one who does: Lizbeth, who works with ChatGPT and has gone from one novel a year to three. It also quotes Jason Hamilton, who runs a YouTube channel on the subject, finds drafting “agonisingly slow”, and hands that part to AI before editing it heavily. He’s frank that “there is something lost there”, and he thinks more is gained. They’ve both made a choice about their own books, and that choice is theirs.

What AI Fiction Tends to Do

It tells the same stories. A Cornell study by Sil Hamilton and David Mimno sampled 20,000 stories from four current models and found that 88.3% of them contained at least one of 11 words, among them Elias, Mara, Elara, lighthouse, clockmaker and librarian. Those words are rare in published fiction. If your draft has acquired a lighthouse keeper called Elias, he didn’t come from you.

It has recognisable faults. For a study at CHI 2025, professional writers edited over a thousand AI-written paragraphs of literary fiction and creative non-fiction, and agreed on seven kinds of trouble: cliché, needless exposition, purple prose, poor sentence structure, lack of specific detail, awkward word choice and slips in tense. They found no clear winner among the three models they tried. Those were 2024 models, and newer ones may do better, but it’s a useful list to have open while you read.

Anyone who reads a lot of this will add a few of their own. The emotion that gets named straight after being shown. The small, quiet realisation that closes each paragraph. The contrast (“it wasn’t anger, it was something older”) and the rule of three. None of these is wrong. People wrote them first, which is how the models learned them. The trouble is how often they come, and how little they cost.

It pulls towards the middle. In a 2024 experiment in Science Advances, 293 people wrote eight-sentence stories, some with AI-suggested ideas. Those ideas raised ratings, most of all for the weakest writers, and the stories that used them also resembled each other more. In a CHI 2025 study, an AI autocomplete tested on 118 people in India and the US moved the Indian participants’ writing towards American styles. Neither study involved novelists or a whole book. They do suggest a direction.

Now the evidence against the comfortable view. A 2025 preprint trained ChatGPT on the complete works of each of 50 award-winning authors, then had it write excerpts of up to 450 words in their style. Readers in a blind test, including writers with MFAs, preferred them to imitations by MFA-trained humans. Detection tools flagged only 3%. So “a machine can’t do style” isn’t a safe thing to believe. But look at what it took: an author’s whole body of work, used to produce pastiche in short pieces. It imitates what you’ve already done.

How a Voice Actually Goes

Your voice is the sum of many small choices, a lot of them refusals: the word you use too often, the sentence that runs on because you like it, the joke at the wrong moment, the thing you won’t say about the dead. Here are two sentences, both written for this page. The second is an imitation, but you’ve read it before.

Gwilym had never once thanked the dog, and he wasn’t going to start now it was dead.

Gwilym stood over the dog’s still form, a wave of grief he hadn’t expected washing over him.

Nothing’s wrong with the second one, and that’s the danger. You’d never accept it as your best, but at the end of a long day, you might accept it as fine. Suggestions don’t usually overrule your style. They wait until you’re tired, then offer something acceptable. One costs nothing. A few thousand and the book has a new author, and because each accepted sentence becomes part of what the model continues from, the drift compounds.

Some of what’s lost is harder to see. Stephen May, a novelist quoted in the report, says the first draft is “like it’s being pulled out of me, and that is a good thing”. Not everyone agrees, but it matches what plenty of writers find: when you can’t get a scene to work and sit with it long enough, the thing you eventually come up with is often the best idea in the book. A generator gets you past that moment, and you won’t know what you’d have found.

If You Use It, What Keeps a Voice

Keep a control sample. Before you start, put aside ten pages you’re sure are yours. Read a page aloud before each session with AI, then read aloud what you’ve just written. Your ear will notice what no tool will.

Ask for questions, not sentences. “Where does this chapter lose tension?” “Which words have I repeated?” One novelist in the report, Fia, explains why this is different: “It’s told me I’ve repeated this, I’m going to go and find the words that I want to use.” The choice stays with you.

Retype, never paste. If you do want a suggestion, read it, close it and write your own version. Your fingers will disagree with it in the right places.

Use it to find what to avoid. Ask for ten ways the scene could go, and treat the list as the ones every other writer’s tool is offering too. Then write the eleventh.

Watch “improve” and “tighten”. Run a page through and compare it with the original, word for word. Note what’s gone. It’s often the odd bits: the repetition you meant, the broken rhythm, the word only your family uses. That’s the part to put back.

Don’t accept anything when you’re tired. Close the assistant at the same time you’d stop trusting your own judgement.

Check the facts. Looking things up was the use novelists minded least, but models make things up with confidence. A crime writer in the report uses it to learn what machinery looks like after decades underwater, and often checks the sources.

Read the terms before you paste a book in. Find out where your text goes, and whether the provider may train on it. A model that runs on your own computer avoids the question.

Some writers give the model their earlier books and a style guide, as Lizbeth does. That may reduce the drift, though what you get is an average of you. She told the researcher she hopes one day to have it write chapters in her style for her to edit, and that the line then becomes “a little bit more fuzzy”. You’ll have to decide how you feel about that.

When Not to Use It at All

On your first book. You don’t have a baseline yet. Those clumsy pages are how you find out what you sound like, and a tool that smooths them removes the evidence. Learn it the slow way once.

On memoir, and on anything where the point is that you said it. A eulogy with good rhythm from somewhere else is a strange thing to have read out.

When the contract or the rules say no. Some publishers and competitions have a policy, and KDP asks about AI content when you publish. If you’re not sure, ask first.

When you’re stuck on the scene that matters. This is when it’s most tempting, and when the answer is most likely to be the one you’d have reached on your second day of thinking, and then rejected.

When you only feel you ought to. Some writers told the report they use it because they felt they had to “keep up”. That’s a poor reason to change how you make something you love. Plenty of finished novels are being written with nothing but a keyboard, and that’s a legitimate way to work. It isn’t going to stop being one.

When you like the work. Writing a novel is hard, and for many of us, it’s also the reason to bother. If drafting is the part you enjoy, nobody can make a case for automating it.

Who Wrote This, and Why You Should Check

We make Shiro, a writing app with an optional AI in it, so we have an interest here, and that’s why every figure above is linked to its source. If you’ve decided you want no AI at all, we have a page on writing without it. You should also know that Scrivener, Dabble and Novlr each state on their own sites that they keep AI out of your writing. Any of them would be a perfectly good place to write a book.

The Questions People Ask

Can AI copy my writing style?

To a degree. Prompted with a sample, models tend to catch the surface. In one 2025 study, a model trained on each author’s complete works produced short excerpts that readers preferred to imitations by trained writers. That was 450 words at most, and it imitates what you’ve already written, not what you’d write next.

Do published novelists use AI?

Some do, but mostly for chores. In a 2025 Cambridge survey of 258 published UK novelists, 67% said they never use it in their work and 22% rarely. The most common use was looking things up (20%). The figures may differ among self-published writers, who were only 10% of that sample.

Will readers be able to tell?

Sometimes, and less often as models improve and writers edit. Detection tools aren’t a safe guide in either direction. Writers in the Cambridge report also worried about being wrongly accused, which shows how little certainty there is.

Do I have to say I used AI?

It depends on the contract and the store. Ask your publisher, read competition rules, and if you publish through Amazon KDP, expect to be asked about AI content when you upload. We have a separate guide to that question.

Last revised November 2026.