Sentence rhythm
Short impacts, long arcs, variation, and pauses shape pace. A grammatically correct edit can still break the rhythm of a scene.
Home › Voice-Preserving AI Editing
Our newest innovation
Your text should become better. Not less yours.
EPOS connects three editing levels with a voice profile built from your manuscript. A suggestion is judged not only by correctness, but also by whether it still belongs in your work.
Voice-preserving AI editing improves language without pushing every manuscript toward the same polished register. EPOS considers sentence rhythm, syntax, word choice, and deliberate punctuation from your own work. You see every suggestion and make the final decision.
The most common objection to AI editing is valid
A sentence comes back technically correct. The repetition is gone, the punctuation is tidy, and the prose flows more smoothly. Yet it no longer feels like yours. The loss rarely comes from one dramatic mistake. It comes from many small edits that all move toward the same average style.
Generic optimization often rewards clarity, regularity, and familiar sentence shapes. Fiction also depends on breaks, pace, omission, resistance, and recurring forms. EPOS does not treat those features as errors by default. It considers the edit in relation to the manuscript.
Documented internal measurements
A meaningful signal, published with its limits
73%Average reduction in the measurable distance from the author’s own writing. Three runs on two chapters, ranging from 58 to 88 percent.
10 to 0Broken form rules per chapter in the documented measurement.
35%Correct author among 30 candidates, compared with 3.3 percent chance.
60German-language authors with two works each in the recognition test.
The profile was built from work one. A 600-word test passage came from the previously unseen second work. The correct author appeared in the top three results in 62 percent of cases.
What this does not prove: The measurements do not guarantee a perfect edit. They are internal, the 73 percent result comes from three runs on two chapters, and an independent multi-author blind study is still pending.
What "voice" actually means, and why AI tends to flatten it
Voice is not one thing. It's the combination of sentence rhythm, syntax, word choice, and punctuation that makes a passage recognisable as yours rather than a competent rewrite of it. Independent research outside EPOS shows why general-purpose AI editing struggles here specifically. A University of California, Berkeley study revised 300 first-person narratives with three frontier language models under three prompt conditions, including an explicit instruction to preserve voice, and measured 13 stylometric markers (Van Nuenen, "Voice Under Revision: Large Language Models and the Normalization of Personal Narrative", arXiv:2604.22142, submitted 2026-04-24, accessed 2026-08-08). The finding: "every model in every condition drifted in the same direction," reducing function words, contractions, and first-person pronouns while increasing vocabulary diversity and word length, regardless of what the prompt asked for. Explicitly asking a model to preserve voice helped, but did not solve the problem: "voice-preserving prompts reduce effect magnitude by 32% but preserve direction," meaning the average drift only shrank, from an effect size of 1.11 to 0.76, it did not disappear. This is the closest independent, non-vendor evidence available for why voice preservation needs to be a deliberate part of an editing tool's design, not an assumption. It's also a useful benchmark for reading EPOS's own numbers above with the right amount of skepticism: an internal 73 percent result is a meaningful signal precisely because independent research shows the underlying problem is real and consistent across models, not because internal numbers alone would otherwise be convincing.
Four layers that make prose recognizably yours
Syntax
Main clauses, subordinate clauses, fragments, and information order all shape the stance of a narrative voice.
Word choice
Several words may be correct. The decision between them is often where a writer becomes recognizable.
Punctuation
Not every mark is only a rule. A period, comma, colon, or omission can change breath and emphasis deliberately.
An example: correct is not always right for the voice
He lied. Of course he lied. The truth never lasted long with him.
The repetition and short sentences carry distrust and hardness.
Of course he lied, because the truth never lasted long with him.
The sentence flows, but the original beat and repeated accusation disappear.
Voice preservation does not mean the original can never change. It may need correction. The question is whether the edit can solve the issue without replacing the effect with a borrowed tone.
You decide how EPOS treats a pattern
This is me
The pattern belongs to your voice and should not be smoothed without a strong reason.
Please avoid
You want EPOS to work deliberately against this habit.
No preference
You leave the decision to the selected editing pass.
Signs an edit is quietly flattening your voice
"Not just X, but Y"
This and similar contrast constructions show up disproportionately often in AI-revised prose, regardless of the original tone of the text.
Stacked descriptive adjectives
Several ornamental adjectives in a row where the original had one precise word.
Sentences converge on the same length
Short, broken sentences get smoothed into uniformly longer ones, and the pace and edge of the original disappear.
Extra connective words
Words like "however," "therefore," or "moreover" appear more often, even where the original left the connection implicit on purpose.
These are signals, not proof. One smooth sentence is not a problem. When several of these patterns cluster on a page, it's worth checking whether the passage still sounds like you, rather than judging each sentence in isolation.
What to look for in any tool that claims to preserve your voice
- Does it build a profile from your own writing, or only apply generic "sound more natural" instructions to everyone?
- Can you mark a specific pattern as intentional, so the tool stops flagging it?
- Does it show you the reasoning behind a suggestion, or only the suggestion itself?
- Does it publish any measurement of its own results, with the method and sample size disclosed, rather than only a marketing claim?
None of these questions has a universally right answer, but a tool that can't answer them at all is asking you to trust it on faith. Vendor claims about preservation rates are common in this category; treat a number without a disclosed method and sample size the same way the research above treats unverified marketing claims, as unconfirmed until shown otherwise.
When AI editing helps your voice, and when it works against it
Mechanical and structural passes
Spelling, punctuation consistency, and flagging a genuine continuity error rarely touch voice, because these edits have little room for stylistic interpretation.
Broad stylistic smoothing
An instruction like "make this flow better," applied without context, is exactly the kind of edit the research above shows drifting every model in the same direction.
This distinction matters when you decide which editing level to run. Copyediting and consistency checks carry little voice risk by nature, a broad stylistic or developmental pass carries more, and it's worth reviewing that kind of suggestion more carefully for exactly this reason.
From voice profile to edited manuscript
Provide enough of your own prose
A preliminary profile is possible from 3,000 words. EPOS recommends 20,000 words for a strong profile.
Select the editing level
Copyediting, line editing, and developmental analysis have different jobs.
Review suggestions in context
Inspect the issue, the reasoning, the effect, and the connection to the wider manuscript.
Remain the author
You decide which change strengthens the work and which distinctive choice stays.
For individual authors and professional teams
Authors
Edit a complete novel
Three editing levels with manuscript context and visible suggestions.
Independent editorsPrepare a second view
Collect patterns and findings, interpret them, and discuss them with the author.
PublishersStructure editorial preparation
Include author voice as a quality criterion in a human-led workflow.
Common questions
Is voice-preserving the same as error-free?
No. Error correction and voice proximity are different goals. A useful suggestion must address the issue and still fit the manuscript.
Can EPOS copy my voice perfectly?
No. The profile describes measurable patterns. It cannot replace your intent or every meaning you attach to a sentence.
Does it work with a short sample?
A preliminary profile is possible from 3,000 words. EPOS recommends 20,000 words for a strong profile.
Are the results independently validated?
Not yet. The published figures are internal measurements. An independent multi-author blind study is still pending.
Can I reject suggestions?
Yes. You keep control over every change to your manuscript.
What does independent research say about AI and writing voice, not just EPOS's own numbers?
A University of California, Berkeley study found that every tested model drifted prose in the same direction regardless of instructions, and that explicitly asking a model to preserve voice reduced but did not eliminate the drift. Full details and citation are above under "What 'voice' actually means, and why AI tends to flatten it".
How can I tell if an edit is flattening my voice?
Common signals include contrast constructions like "not just X, but Y," stacked descriptive adjectives, and sentences converging on the same length. Examples are above under "Signs an edit is quietly flattening your voice".
What should I look for when comparing tools that claim to preserve voice?
Whether the tool builds a profile from your own writing rather than generic instructions, whether you can mark a pattern as intentional, and whether it discloses its own measurement method. Full list above under "What to look for in any tool that claims to preserve your voice".
Is voice preservation only relevant for fiction?
The four layers this page describes (sentence rhythm, syntax, word choice, punctuation) apply to any first-person or strongly authored prose. This page and its documented measurements focus specifically on fiction manuscripts.
What does this cost compared to a traditional human editor?
Voice-preserving editing is part of the EPOS plan, not billed separately. What editing costs overall is covered in Editing Costs 2026.
Does EPOS train its own models on my manuscript to learn my voice?
EPOS-AI does not use manuscripts for its own model training. The voice profile is built from your text for your own editing sessions, not used to train a general model.
Read the evidence and continue
Improve the manuscript without forcing it toward someone else’s ideal.
7 days free. Cancel in one click. A card is required to start the trial.
7 Days FreeNew focus areas: Voice-preserving AI editing · Pro Editing for publishers, agencies, and editors