You know the wall. You paste a chapter of your own writing in. You ask for a rewrite, or a continuation, and you add the magic words: keep my voice. What comes back is competent. Sometimes it is better than what you gave it. And it is not yours.
I built bookmoth because every tool I tried either didn’t nail my voice, or didn’t even consider that I’d want to keep it. Chapter one would sound fine, because the model had just read my samples. By chapter three it was writing like itself again, and nothing had noticed. A new paper from Berkeley researcher Tom van Nuenen has now measured that drift properly, and the finding is worse than most writers assume. No wording of the prompt fixes it. That is not a reason to give up on the tools. It is a reason to be picky about what they do after the prompt.
First, a story that shows the problem from the other side. In April, fantasy author Lena McDonald published Darkhollow Academy: Year 2. Inside the finished book, readers found this sentence sitting in the prose:
“I’ve rewritten the passage to align more with J. Bree’s style, which features more tension, gritty undertones, and raw emotional subtext beneath the supernatural elements.”
She had pasted her editing prompt into the manuscript and missed it on the copy-edit. The internet did the rest, and the story ran as another author caught using AI without saying so.
The bit I keep thinking about is that the prompt worked. She asked a model to write like another author, and it did, well enough that the only reason anyone noticed was the clipboard slip. Models are very good at sounding like a named writer for a page. What they are bad at is staying that writer for a book, and nothing in the prompt box changes that.
Why does AI writing flatten your voice even when you tell it not to?
Because the model has a voice of its own, and it is heavier than yours. Claude, ChatGPT and Gemini are post-trained on enormous amounts of edited prose, and that training rewards what professional editors like: cleaner sentences, more standard vocabulary, tidier punctuation, a little distance from raw first-person narration. That preference becomes the model’s centre of gravity.
When you ask it to write in your voice, you are asking it to hold itself away from that centre. It can, for a sentence or two. Then every generated sentence is pulled a little further back toward home. By the third or fourth paragraph the voice you started with is gone, and you did nothing wrong. The instruction is still sitting there in the prompt. It just is not strong enough to win a tug of war a thousand times in a row.
If you have seen the complaints about Claude Opus 4.7 in writer forums, “memo voice” and “reaches for bullet points” and “feels like an email,” that is the same pull, just louder. It is not user error and it is not bad prompting. It is what a prompt is.
What did van Nuenen's study actually prove about AI revision?
That every revision prompt moves prose in the same direction, and the “preserve my voice” prompt only moves it a bit less far.
The setup is simple enough to describe in a paragraph. Van Nuenen took 300 first-person narratives and ran each one through three frontier models under three instructions: improve this, rewrite this, and revise this while preserving the original voice. Then he measured thirteen stylometric markers in the input and the output. Function words, contractions, first-person pronouns, vocabulary spread, word length, punctuation, emotion words, and a few more. They are the same kind of fingerprint that let Claude Opus 4.7 identify a journalist from 125 unpublished words earlier this year, measured directly instead of guessed.
Every model, under every instruction, drifted the same way. Fewer function words. Fewer contractions. Fewer first-person pronouns. Longer words, wider vocabulary, fancier punctuation. Embedded narration became distanced narration. The preserve-voice prompt reduced the size of the shift and did nothing to its direction.
Put plainly: every AI revision makes your prose more polite, more formal, and slightly less like the person who started the sentence. Even the one you told not to.
[Source: van Nuenen, “Voice Under Revision: Large Language Models and the Normalization of Personal Narrative,” arxiv 2604.22142, April 2026.]
Can Sudowrite, NovelCrafter, Claude, or ChatGPT actually preserve your voice?
Not across a novel, and for the same reason in each case. Writers switch between these tools looking for the one that does not drift, so it is worth saying what each one actually does with your voice.
Sudowrite gives you a Codex for your world and characters, style descriptions, and sample passages. It is a much better setup than a bare chat window. All of it still goes to the model as text it is asked to follow, and its Muse model drifts inside a chapter like everything else.
NovelCrafter lets you bring your own key and pick your model, plus a Codex of its own. Picking a model helps a little, because some drift less than others. It does not change what happens after the prompt.
Claude or ChatGPT directly, with custom instructions, projects, or pasted samples, is the exact setup the study tested. It fails the way the study says it fails.
I am not knocking any of them. Sudowrite and NovelCrafter are strong on planning and world-building, and if you are hunting for a Sudowrite alternative or a NovelCrafter alternative, the honest comparison is on those pages. On the single question of whether chapter twenty sounds like chapter one, they all hit the same ceiling, because they all stop at the prompt.
What kind of AI writing tool actually preserves your voice across a novel?
One that treats your voice as something to measure and check, not something to describe and hope. I call that constraint-based, as opposed to prompt-based, and the difference is in what happens before and after the model writes.
In a prompt-based tool, you describe your voice, or paste some of it, and that goes to the model alongside the request. The description anchors the first sentence or two, then the pull takes over. That is the whole mechanism, and it is what the study measured.
In a constraint-based tool, your writing is analysed before any drafting starts. The rhythm of your sentences, how long they run and how much they vary, how much of a page is dialogue, how you handle interiority, what you do with adverbs and paragraph breaks. That becomes a profile built from your prose rather than from adjectives, and every chapter is drafted under it. Then the finished chapter is measured against your samples on those same dimensions, and if it has drifted, it is corrected before you see it. The rule is not something the model promises to obey. It is something the draft has to pass.
That is how bookmoth works. It also runs on Claude Opus 4.6 rather than 4.7, because in my testing 4.6 is easier to move off its own voice, and a tool built to hold your voice should start from the model with the lighter pull. Nothing about that is magic. It is measuring twice, on every chapter, without you having to think about it.
How does the constraint-based architecture work in practice?
You give bookmoth your own prose to start with. A few chapters of a previous book, or of this one, or essays, anything that sounds like you. Five thousand words is enough to begin; more is better. It reads them and builds your Writing Profile, which you can open and read. Writers tend to have a strong reaction to this part. It is the first time most of them have seen their own habits written down.
From then on, every chapter is drafted under that profile, and every draft is checked against it before it lands in your manuscript. You never re-explain yourself, because the profile, the brief, the plan, and the earlier chapters are all still in the room.
The reason this matters is length. A short story is five thousand words and a novel is ninety thousand. A drift you would never notice in a single scene compounds over forty chapters until the book reads like it was written by someone else. The check on every chapter is what stops the compounding.
The van Nuenen paper is the measurement of the problem I have spent the last year building against. If you have felt the drift in Sudowrite, NovelCrafter, or Claude and assumed you were prompting badly, you were not. The prompt was never going to be enough on its own. Something has to read the chapter afterwards and know how you write.