An AI edit that keeps your voice.

There is a moment every author knows who has ever run a manuscript through an AI. The text comes back. The commas are right. The repetitions are gone. The paragraph you spent three evenings on now reads more smoothly than before. Everything is better.

And something is missing.

You read the passage again. You cannot find a fault. You cannot even point to the sentence. But the text is no longer entirely yours.

Why this is no accident

A language model has read millions of texts. Out of that reading it has formed a taste, and that taste is the average of everything it has ever seen. It prefers middling sentence lengths. It likes variation at the start of a sentence. It smooths repetitions, even when a repetition was meant as an echo. It replaces the bare word with the more precise one, even when the bareness was the character.

Every one of those interventions is defensible on its own. An editor could justify it. The trouble only appears in the sum: they all point the same way, towards the middle. And the middle is precisely the place where no author is distinguishable from any other.

Errors are visible. Drift is not. On one page it cannot be seen. Across 500 pages you have a different book.

We were asking the wrong question

For a long time it was: how do we stop the AI from changing the style?

The answer to that is always the same, and always poor. You make the AI more cautious. It suggests less. It stops daring. In the end you have an edit that finds nothing and an author who switches it off.

The right question is a different one: what is it that makes an author recognisable at all?

If that can be answered, the rest is craft. Then every suggestion can be held against it before the author ever sees it, and the caution no longer sits in the instruction. It sits in the check.

How you measure a voice

Not by what someone writes about. That changes from chapter to chapter.

By how they do it. And specifically by the things no author controls deliberately: the rhythm with which sentences follow one another. The build of the sentences themselves, the balance of main and subordinate clauses. Word choice in the moments where several correct words exist. Punctuation, especially where it is not a rule but a decision. And the way characters speak, and how their speech is set into the surrounding text.

That is the layer on which a text sounds like someone. And because it is unconscious, it is stable. An author can change subject, genre, point of view. The rhythm stays.

From those observations a voice profile emerges: a description of what is normal for this one author. Not what they do correctly. What they usually do.

10 times better than chance

A measurement like that could be pure noise. You can count anything, and numbers always look like insight. So we put it through a test that cannot be talked around.

60 authors, two works from each. On 30 of them we tuned the method, on the other 30 we measured. And we did it like this: the voice profile was built from one work, the test passage came from the second, which the system had never seen.

Then it was given 600 words from that unknown book and 30 candidates to choose from, and had to answer: who wrote this?

60 authors, public corpus.

That is the point where a hunch becomes an instrument.

The error that cost us 14 percentage points

That figure started at 21 %, not 35 %. It rose because we found a mistake of our own.

60 authors, public corpus.

The weighting of the feature groups came from an earlier experiment. But that experiment had answered a different question from the one that matters. It had measured which groups change when an AI smooths a text. Not what makes an author recognisable. Two questions that sound alike and have different answers.

When we measured the right one, the ranking turned almost upside down. The group we had given by far the greatest weight was the weakest of them all.

We replaced the weighting, and not with one we had gone looking for in the data. With the simplest assumption available: every feature counts the same. It wins on both halves of the experiment, the half used for tuning and the half never seen. That is exactly what separates a result from an arrangement.

Zero is unreachable

We had a measure meant to say how close a suggestion sits to the author. It counted down to zero. Zero meant perfect. And zero is unreachable.

An author's own sentences deviate from their own average too. They have to. An average is not a sentence.

An average is a statement about many sentences, and none of them. To count towards zero is to demand that the edit write more evenly than the author does, and to penalise it for failing.

We now count towards the floor the author reaches themselves: towards the distance their own sentences keep from their own mean, at exactly this text length.

Internal measurement, three runs.

How we measured

There is a second observation we find almost more important. An author's hard habits, the marks and forms they never use, used to come back through the suggestions in around 10 cases per chapter. Today in 0.

Internal measurement, three runs.

The same data as before. Only without the fallacy.

Measuring is not enough

This is where most systems stop, and it is where it starts to get interesting.

For a long time our method measured every deviation cleanly and delivered the suggestion anyway. That is an editor marking the fault and handing back an unchanged page. The measurement was there. It did nothing.

Today every suggestion that falls out of range gets a second attempt. Not with the instruction write it better, but with the specific deviation in the brief: here is what you did, and here is what this author does in this place.

Three rules protect it. The original finding must remain fixed, or the second version repairs the form and leaves the fault standing. The closer of the two versions wins; a step backwards is discarded. And breaking one of the author's hard habits weighs more than a better number, because it is not a matter of taste.

It costs you nothing extra. It happens before you see the suggestion.

Internal measurement, three runs.

In the end you decide

An instrument that decides alone would only be a different kind of paternalism.

So EPOS shows you what it found in you. The things you do noticeably often, with a figure beside them and an example from your own manuscript. And you decide on each one:

That is me. Then it is protected and never corrected away.
Please avoid. Then the edit works against it deliberately.
No opinion. Then it stays as it is.

An author who knows they begin every second paragraph with And has a decision to make. An author who does not know it has a habit. The difference between the two is half the craft.

What it is explicitly not

It is not a ghostwriter. EPOS does not write for you.

It does not imitate anyone else's style either. You cannot set it to make your text sound like somebody. The profile is built solely from your own manuscript, and it applies only to you.

And it does not replace editing. It makes sure the editing you get sounds like you and not like an editor with a taste of their own.

What we do not know yet

Two things, and we would rather say them ourselves.

On very short passages the recognition weakens. Below a few lines there is too little making in the text to recognise anyone by. For a chapter this does not matter. For a single sentence it does.

And the final proof is still outstanding. We do not yet have an independent study in which several authors blindly judge which of two versions sounds like them. Until then we say what we can demonstrate: that we measurably prevent drift in form and register, and that we check every rewrite inside the surrounding text where it will actually sit.

That is less than we could claim. It is everything we can show.

When it starts working

A provisional profile forms at around 3,000 words. At around 20,000 it becomes good. English and German are handled separately, each with its own instrument, because a translation of one does not work for the other.

You will find the switch in the editor and in Pro Editing. A line beneath it tells you how well EPOS already knows the way you write.

It took us a long time to arrive at the right question. The answer came surprisingly fast.

We are not done with this

Everything above is a beginning.

A publisher can edit a manuscript, typeset it, print it and market it. An AI can find errors, check structure, make suggestions. Both are craft, and craft can be learned, bought and replaced.

Not the voice. It is the one thing about a book that nobody else can manufacture.

Two authors can tell the same story, the same character, the same scene, the same content in the same sentence. You get two different books. What separates them is not the plot. It is the way a person puts the world into sentences, and that is something they have written their way into over years.

So for us this is not one feature among many, now finished and ticked off a list. It is the question on which it will be decided whether AI helps writing or harms it. A tool that pulls every text a little way in the same direction makes literature poorer, however good each individual suggestion may be. A tool that knows the author makes that author clearer.

What we are working on next

The independent blind study with several authors. It is the one proof we still lack, and we will publish its result even if it turns out to be inconvenient.

Short passages. A single sentence carries less making than a chapter, and the recognition is weaker there today. We know why. We do not yet know how far it can be improved.

More languages. Each needs its own instrument, because translating the existing one would be precisely the mistake we are trying to avoid: a foreign yardstick held against a text it does not know.

And whatever you write to us. The two largest improvements of this year did not come out of a planning meeting. They came from two emails by authors who looked closely and asked politely. Those emails are worth more to us than any market research.