Does AI Editing Preserve an Author's Voice?

Summary

This page documents two measurements on a question that comes up regularly in discussions of AI editing and is rarely backed by evidence: how much of what makes an author recognizable survives a machine intervention in their text?

We split the question into two separate measurements because it concerns two different things.

The first measures the distance to the author's own writing: how far a revised chapter moves from how its author normally writes. It comes from a single manuscript project and is therefore a single case.

The second measures recognizability: whether an author can be identified from a style profile at all. It comes from a public literary corpus with 60 authors and is the more robust of the two.

The core figures:

An independent blind study with multiple authors is still pending. We are not conducting it ourselves because we have a stake in the outcome. Anyone who wants to run it will find the full parameters below.

Why this question has to be asked at all

Language models are trained on probability. They suggest what commonly occurs in similar contexts. That is why they write fluently, and the same reason they pull toward the average.

In practice that means: a model smooths sentence lengths toward an average value, replaces rare turns of phrase with common ones, resolves repetitions that were placed deliberately, and evens out the rhythm between paragraphs. Each of these changes is an improvement on its own. They are exactly the changes an edit is supposed to make.

The problem does not arise in one paragraph. It arises across thirty chapters.

An author who has one chapter revised sees a cleaner version of their text. An author who has an entire manuscript revised chapter by chapter ends up with a book that sounds equally good everywhere and distinctly like itself nowhere. The effect is invisible from chapter to chapter and unmistakable across a whole book.

That is exactly why it is not enough to instruct a model to preserve the voice. An instruction shifts a probability distribution, it does not cancel it. We are not alone in this assessment, it matches independent research cited further below.

So the question is not whether a model writes better. The question is: how much of what makes an author recognizable survives the intervention, and can that be measured instead of claimed?

What "voice" means in this context, and what it does not

The term voice is used broadly in literary criticism. It covers stance, perspective, choice of subject, and worldview. None of that is measurable, and we do not claim to measure it.

What is measurable are observable features of the text surface: the distribution of sentence lengths and their spread, paragraph lengths, vocabulary diversity, the distribution of function words, punctuation habits, sentence openings, the share of direct speech, recurring patterns.

These features are not the voice. They are its imprint on the surface.

The difference matters, and marketing copy regularly blurs it. A tool that measures stylistic shift cannot determine whether a text has kept its artistic identity. It can determine whether it has moved, in measurable features, away from what the same author otherwise writes. Everything beyond that is a judgment, and a judgment needs a reader.

We consider the second claim useful and the first one dubious.

Measurement 1: the distance to the author's own writing

Data basis

A novel manuscript from EPOS's own house account. No customer account, no third-party work, no third-party consent required. Two chapters, three independent runs.

Method

For each tested chapter, a style profile was built from the rest of the manuscript, that is, exclusively from text the given run never saw. This separation is the core of the method: a profile built from the same chapter it later evaluates measures itself.

The chapter was then edited once with the voice profile and once without. We measured the stylistic distance between the result and the author's original writing, in the same features each time.

The metric "closed gap" denotes what share of the distance created by an edit without a profile is recovered by the profile. A value of 100 percent would mean: the result is no longer distinguishable from the original writing. A value of 0 percent would mean: the profile has no effect.

Result

The third row measures something different from the first two. Form rules are explicit instructions from the author about the form of their text, for instance about tense or the handling of direct speech. Without a profile, an average of ten such instructions were violated per chapter, with a profile, none.

How to read these figures

The range matters more than the average. 58 to 88 percent across three runs on the same two chapters means the result varies considerably. An author who has a chapter revised does not reliably get 73 percent. They get something within this range, and which end they land on is decided by the individual run.

And the sample size cannot carry a general claim. One work, one author, two chapters, three runs. Anyone who turns this into a performance claim for all manuscripts is stretching the data considerably. We do not do that, and we ask that it not be done when citing this either.

Measurement 2: the recognition test

This measurement is the more robust of the two, because it was carried out on works by other authors and because its sample size is two orders of magnitude larger.

Data basis

60 German-language authors with two works each, from a public literary corpus released for reuse. No customer data, no manuscripts from active projects.

Method

A style profile was built from each author's first work. It was tested against a 600-word passage from the second work, entirely unseen by the profile. The task: name the correct author among 30 candidates.

The separation between the profile work and the test work is the decisive part of the setup. It tests whether a profile recognizes an author beyond a single work or merely recognizes one particular book. These are two very different capabilities, and only the first one is useful for editing.

Result

How to read these figures

35 percent is ten times chance. With 30 candidates, chance expectation is 3.3 percent. A method that is right ten times more often than guessing is demonstrably measuring something.

35 percent is, at the same time, far from a reliable identification. In two out of three cases, the first guess is wrong. Anyone who takes this figure as proof that a style profile reliably identifies an author is reading it wrong.

Both statements belong together, and we ask that they not be separated when cited.

The 62 percent figure for the top three is the more meaningful one for practical use. An edit does not need to identify an author. It needs to recognize which direction their writing leans. For that task, a placement among the top three is already useful.

The ablation is the methodologically most interesting figure. Remove the strongest feature block from the method, and the hit rate falls from 35 to 21 percent. So it does not collapse, it drops significantly. That means the effect depends on an identifiable group of features, but not exclusively on it. No single feature carries the method, and none brings it down either.

What independent research says about this

Three papers from 2025 and 2026 touch on the same question. We were not involved in any of them and have no relationship with any of the groups.

Van Nuenen, "Voice Under Revision", UC Berkeley

arXiv:2604.22142, submitted April 24, 2026.

300 personal texts, three top-tier models, 13 style markers. The finding: rewriting by a language model systematically normalizes style. Voice-preserving instructions reduce the shift but do not reverse its direction.

This supports our starting thesis and, at the same time, qualifies our own result. It supports it because it independently shows that prompting alone is not enough. It qualifies it because the same logic applies to a style profile: a profile shifts the result, it does not cancel the effect. Our own 73 percent is not evidence that the shift disappears, but that it becomes smaller.

Maier, Zaiss, and Bayer, FAU Erlangen

arXiv:2605.02620, May 4, 2026.

The paper reports that GPT-5.5 and Claude Opus 4.7, with targeted post-editing, close 71 to 75 percent of the style gap to the same-author ceiling.

This range sits right around our own value of 73 percent. We state this explicitly instead of glossing over it, because it is the most obvious critical question for our measurement: does a voice profile achieve anything more than a carefully used standard model?

Two differences matter to us, and both remain unproven until someone tests them.

The cited paper measures on individual texts, our measurement on chapters within a manuscript. Whether the effect stays stable across an entire book is a different question from the one about a single text.

And the cited paper measures with targeted post-editing per text, our measurement without author intervention between chapters. The effort required on the user's side therefore differs considerably.

Whether this difference justifies the effort of a voice profile is not settled by the available data. We consider it likely and have not proven it.

Wang, Tripto, Park, Li, and Zhou, "Catch Me If You Can? Not Yet"

arXiv:2509.14543, 2025.

The paper deals with the detectability of machine-generated text. It is indirectly relevant to our question: it shows how hard machine editing is to prove at the text-surface level.

For authors, this implies a point that is often misunderstood. A publisher generally cannot determine whether a text was edited with AI. What stands out is something else: text that sounds equally good everywhere and distinctly like itself nowhere. That comes from smoothing, not from the use of AI as such.

Limits of these measurements

This section deliberately does not appear at the end, but before the practical conclusions.

What follows from this in practice

For authors who revise with AI, three things follow from both measurements.

Prompting is not enough. The Berkeley paper independently shows that voice-preserving instructions reduce the shift but do not reverse its direction. Telling a model to preserve its style gets you a smaller shift, not none at all.

A profile built from the author's own work shifts the result measurably. By how much depends on the run. The form rules are the most reliable part of this: ten violations per chapter down to zero is a clear difference.

And control stays with the author. No method replaces reading through the suggestions. A value of 73 percent does not mean that 27 percent of the changes are wrong. It means the distance does not go to zero, and that someone has to decide which change belongs in the work.

Reproducibility

Anyone who wants to reproduce the second measurement needs:

A corpus with at least two works per author. Our basis: 60 German-language authors, two works each, public and released for reuse.

A strict separation between the profile work and the test work. The profile must not see any text from the work it is tested against.

A passage of fixed length. Our choice: 600 words.

A candidate pool of fixed size. Our choice: 30, corresponding to a chance expectation of 3.3 percent.

An ablation test. Without it, there is no way to tell whether a result depends on a single feature.

We provide the full feature list and its weighting on request. Anyone who runs the measurement independently will hear no objection from us to the result, whichever way it turns out.

What an independent blind study would need to answer the open question: multiple authors, multiple genres, evaluation by readers who do not know which version was edited, and a comparison group using a standard model without a profile. Until that study exists, our figures stand as this: they show an effect, they do not measure it in general.

Citation

EPOS-AI Research (2026): Does AI Editing Preserve an Author's Voice? Method, internal benchmark, and recognition test. Lucerne, Switzerland. epos-ai.ch/en/research/author-voice-ai-editing-2026.html

The tables on this page may be used with attribution to the source "EPOS-AI Research".

For questions and for the full parameters: info@epos-ai.ch

Status and changes

August 9, 2026: First publication.

This page will be updated when new measurements are available. Changes are noted here, not incorporated silently.

Research conducted by EPOS-AI, Lucerne, Switzerland. EPOS-AI is a working environment for authors working on entire manuscripts.