How Proselon attributes authorship

Last updated 16 August 2026.

Proselon keeps a record of who wrote each character of your manuscript. You can turn it on in any page and see your own words in your own ink and the co-writer’s in another, and you can ask any folder or the whole book for its figures. This page is the specification of that record. It exists because a writer may one day have to describe their contribution to their own book to a publisher or a registrar, and a number nobody can interrogate is worth very little when they do.

The record

Every page carries a list of runs. A run is a hand and a length, the lengths add up to the length of the page, and the three hands are these.

Human: The text arrived through the Proselon editor, typed or pasted by the person at the keyboard.
AI: The text arrived from outside the editor while a writing engine was working in the book.
Unknown: The text arrived by neither route, or it was already there when Proselon began watching. Manuscripts imported from elsewhere begin wholly Unknown, and stay that way until they are edited.

Lengths are counted in grapheme clusters rather than in bytes or code units, so an emoji and an accented letter each count once, and the count means the same thing on every machine.

The three hands are observations. Nothing in Proselon reads your prose and forms an opinion about who wrote it. There is no detector and no stylistic model behind any of it, and no sentence is assigned a hand because of how it sounds. The record states what the app watched happen, which is why it can be relied on and also why it says Unknown as often as it does.

What the record measures

Copyright asks who composed a passage. It does not ask whose fingers were on the keys, and the two questions have different answers more often than one might expect. The United States Copyright Office puts it this way in its 2025 report on copyrightability: where a person modifies material a machine generated, the copyright reaches the material the human author contributed but would not extend to the underlying AI-generated content itself. Korea’s guidance, quoted in the same report, states the rule more bluntly still, that where a person has made additions or deletions, only the part that underwent the change is copyrightable.

Proselon’s run model is that sentence expressed as data. An edit moves the hands of the graphemes it changed and leaves every other grapheme where it was, which has two consequences worth stating outright.

Things that do not make a passage yours

Formatting it. Italicising a sentence, setting it in bold, putting quotation marks around it or wrapping it in a link claims the markers you typed and nothing else. Federal regulation places mere variations of typographic ornamentation, lettering or coloring outside copyright altogether, so there is no quantity of italics that would tip the balance. The Copyright Office reaches the same conclusion from the other side, treating formatting done by a machine as too slight to need disclosing at all.
Renumbering a list it sits in. Adding an item to a numbered list gives every item beneath it a new number, and splitting one list into two starts the second half at one again. Those digits are an index rather than anybody’s writing, so each item keeps the hand it had. The item you typed is yours, its own number included.
Moving it. Dragging a paragraph up the page carries its record with it. Where the move adds a blank line that was not there before, that line is yours, because it is new.
Copying and pasting it. A passage duplicated inside a page keeps the hand of the passage it was duplicated from, whoever pressed the keys.
Retyping it word for word. Deleting a sentence and typing the same sentence back does not originate it, in the same way that retyping a page of somebody else’s novel does not make the page yours.

Things that do

Words you compose are yours from the first keystroke. So is a rewrite, and so is a rephrase, down to the parts of it that changed.

Take a line a co-writer produced, I can afford to allow myself to have a feeling here, and suppose you delete it and type Here, I can afford to allow myself to feel. Proselon records Here, and feel as yours and the run between them as the machine’s, because those words are the ones that survived. What you also did was move the adverbial to the front and turn a noun phrase into a verb, and that is authorship of a kind the record does not score. The section below says why.

Why arrangement is not scored

Deciding where a sentence should open, which paragraph follows which, and what order a scene runs in is your authorship even where the words themselves came from a machine. Proselon does not put a number on it, and that is a decision rather than an omission.

Two different doctrines are at work, and they operate on different units. One is compilation, which the Copyright Act defines as the collection and assembling of preexisting material arranged so that the result is original. That is what the Copyright Office protected in Zarya of the Dawn, where the author’s placement and arrangement of the images and text on each page was held copyrightable even though the images themselves were not. It works on whole assembled things, which in a novel means paragraphs, scenes and chapters, and it needs more than a handful of them before the Office treats the arrangement as more than trivial.

The other is modification, which is the route a rephrase travels. Word order inside a sentence is not the assembling of preexisting material, so a rephrase is judged the way an editorial revision is judged, by whether the resulting work as a whole carries enough original authorship. That question has an answer for a book. It has no answer for a character, and a per-character record can only ever offer a per-character answer.

So a fourth colour meaning machine-drafted, human-revised would put a figure on the page that no doctrine asks for, and it would spoil the one thing these figures do well. Asked what to exclude from a registration today, the record answers without ambiguity: everything not marked Human. A revised bucket would land squarely between the two answers, and deciding which side it belonged on would fall to the writer at exactly the moment they could least afford to guess.

What the record is for is the question the Copyright Office keeps returning to. Registrations fail when the Office cannot tell whose expression is whose. It refused Théâtre D’opéra Spatial not for finding the human’s edits too slight, which it expressly declined to decide, but because the machine-generated material was never disclaimed, and it told the applicant he could file again once he disclaimed it and explained his own contribution. Identifying the machine’s material is what Proselon does exactly. Explaining what you contributed is prose you write, and the arrangement belongs in it.

Elsewhere the line falls in a place that favours you further. The Court of Justice of the European Union has held that as few as eleven words may be protected, since it is through the choice, sequence and combination of those words that an author expresses creativity. Even there the protection runs to the expression as a whole rather than to a scored fraction of it.

Where the line is drawn

Two numbers decide what counts as text that survived an edit rather than text that happens to look similar.

Eight characters is the most that may be added around a passage before the edit stops being decoration and becomes a rewrite. Every emphasis pair a manuscript uses fits inside it, and a sentence of your own does not.
Twelve characters is the shortest unbroken run that is read as having survived. Below that, English hands out matches for nothing — two rewritten paragraphs will share and the and she said without either author having kept a thing, and treating those as survivals would return your own rewrite speckled with somebody else’s colour.

A shorter passage is not lost. It simply is not enough on its own to overcome the reading that you changed this region, which is the reading that credits you with your own edit.

What the figures count

The percentages are computed over characters, because that is the unit the record is kept in and it is exact. The counts beside them are words, because a novelist thinks in words, and a word is assigned whole to whichever hand wrote most of its characters.

Where a word is split evenly, it never counts as yours. That rule is deliberate and it runs one way only. The Human figure is the one a writer would carry into a registration application or weigh against a certification, so a tiebreak that could round it upward is the one imprecision this record must not have. Rounding the other way costs you nothing you cannot see and correct.

Front matter and author notes are left out of the figures. They are not part of the published work, and the numbers are meant to describe the manuscript. Notes are still coloured on screen, so you can always see which of them you wrote.

Every markdown page the file explorer shows as book content is in scope, at any depth. Worldbuilding, plot documents, character sheets and fragments are all counted, not only the manuscript folder.

What it cannot see

The record is built by watching a book, so its reach ends where that watching does. These are the limits, stated because a specification that only lists its strengths is not one.

Text pasted from another page, from another book, or from outside Proselon reads as newly arrived and takes the hand of whoever pasted it. A page is compared against its own previous state, and nothing else.
Carrying a paragraph across many thousands of lines of a single very large page can leave the alignment unable to identify where it went, and the paragraph may then read as newly written. Reordering neighbouring paragraphs is unaffected, which is what reordering normally means.
A line shorter than the twelve-character threshold is not tracked across a move, so a one-word line of dialogue may change hands when it is carried somewhere else.
Editing a book outside Proselon produces text of Unknown origin, which is what Unknown is for.

Where the record lives

Inside your book, beside your manuscript, in plain files. It travels with the folder when you move it; your own backups already cover it. It is never uploaded to us. Nothing about your books reaches Proselon’s servers at all, which the Privacy Policy sets out in full.

Using it

Where a work contains more than a trivial quantity of machine-generated material, the Copyright Office asks an applicant to disclose it and to describe what the human author contributed. The Authors Guild’s Human Authored certification asks a related question, whether the text of the book was written by a person, allowing only a very small amount generated with a machine’s help. Proselon’s figures are meant to be usable for both, and they are calculated to err against overstating your share.

They remain a record of what happened in your book rather than a legal conclusion about it. Whether a given contribution meets the standard for protection is decided case by case, and if a registration or a certification turns on the answer, that is a conversation to have with a lawyer.

Sources

Questions

If the record says something about your book that you cannot account for, write to hello@proselon.com and describe what you did and what it reported. That is a defect worth knowing about.