OneThe problem is not detection
Publishing is spending 2026 trying to work out whether it can catch machine authorship. It is spending almost nothing on how to declare it. That is the wrong way round.
Book deals now collapse over suspected AI use. Publishers cancel releases on allegation. Detection vendors and publishers argue in public about false positives. Agents describe their job as police work. Certification schemes have appeared that attest a book is entirely human-made, and they rest on an honour system.
Every one of those fights is about the same binary: was this book touched by a machine, yes or no. That line was already blurred the first time a novelist ran a manuscript through a grammar checker, and it blurs further every month. A standard built on a line nobody can locate will not hold.
There is an older answer, and the film industry has been running it for ninety years. Do not certify purity. Itemize contribution. Publish a list of who did what and let the reader read it.
A short, fixed vocabulary for stating how a book was made — five credits that can be printed on a copyright page, shown on a retail listing, and carried in book metadata. Nothing here is a purity test, a seal of approval, or a warning label.
It is deliberately compatible with the disclosure work already published by EDItEUR for ONIX metadata, and with the human-authorship certification schemes now in the market. It competes with neither. See section Five.
TwoThe five credits
A complete disclosure is five lines. Any line that does not apply is either omitted or marked none. A book written entirely by one person, unaided, lists that person once and writes none against Models.
Concept
Whose idea the book was. The person who decided this book should exist, chose its subject, and set its argument or story.
Credit: the originating person or people. Note: where a work is commissioned, credit the originator, not the commissioner — the same convention film uses for Story by.
Direction
Who chose the tools and steered them. In a book made with generative models, this is the role that carries the authorship: the person who selected the models, wrote the instructions, judged the output, and decided what survived.
Credit: the human being who did the steering. Always a person, never a tool. Why it matters: existing metadata describes AI as something that assisted a contributor. That frames the model as a pencil. If a model produced prose, someone directed it, and that person is accountable for the result.
Models
Which generative models were used, by name and by the date they were used. Not "AI." Not "an unnamed generative model." The name, as a film names its cast.
Credit: model family and month, e.g.
Claude Opus 5 (Aug 2026). Threshold: disclose anything beyond
de minimis use. Ideation, research, spell-checking, grammar-checking and
proof-reading do not require disclosure. Generated prose, generated images, generated
translation and synthesized narration do.
If none: write none. That is the whole point of having the line.
Editing
Who read the manuscript and shaped it — structural editing, line editing, and the judgement calls about what the book became.
Credit: people, and any model used editorially, named as above. Copy-editing and proofreading may be listed but are not required.
Art direction
Who is responsible for the cover and interior images, and how those images were made.
Credit: the responsible person, plus any model used to generate or materially alter images, named and dated. Include photography that has been generatively extended, replaced or composited — the reader's assumption about a photograph is stronger than about prose, so the disclosure matters more.
Rules of the standard
- People are named. Tools are named. Nothing is anonymous. A disclosure that says only "AI was used" is not a disclosure.
- Silence is not compliance. Omitting the Models line is not the same as writing none. Only the explicit word carries a claim.
- The credit is fixed to the edition. Like a film credit, it describes one release with one date. A revised edition gets revised credits.
- Disclose beyond de minimis. The threshold is borrowed verbatim from EDItEUR's guidance so that the two can be filed together.
- No ranking, no seal, no grade. The standard makes no claim about whether a book is good, or whether AI use is acceptable. It states facts and stops.
ThreeWhat it looks like on the page
Three worked examples. This is a copyright-page block — roughly the size of the Library of Congress data already printed in most books.
A novel written by one person, unaided
- Concept
- Helen Okonjo
- Direction
- Helen Okonjo
- Models
- None
- Editing
- Marcus Reed (structural), Priya Nair (line)
- Art direction
- Sam Whitlock — photography by Ana Duarte
Note what this author gets that she cannot get today: a positive, specific, checkable statement, in a fixed place readers know to look. Not a sticker awarded by a third party — her own signed line.
A non-fiction book drafted with a model
- Concept
- Joe Kuefler
- Direction
- Joe Kuefler
- Models
- Claude Opus 5 (Jun–Aug 2026) — drafting and revision from author outlines and sources
- Editing
- Joe Kuefler; Claude Opus 5 (Aug 2026) — structural review
- Art direction
- Joe Kuefler — cover image generated with GPT Image 2 (Aug 2026)
A translated edition with synthesized narration
- Concept
- Jorge Herrera
- Direction
- Jorge Herrera; translation directed by Lucia Ferrar
- Models
- Translation from Spanish: DeepL Pro (Mar 2026). Narration: synthesized voice “Bryan Lee Jr.”, not a replica of any real performer
- Editing
- Lucia Ferrar — translation review
- Art direction
- Not applicable to this edition
Where a synthesized voice is trained on a specific performer, name the performer. ONIX already provides for this and calls it an authorized voice replica.
FourThe label
A mark for the back cover, the copyright page, or a retail listing. It does not say the book is clean. It says the credits are published and the reader can go and read them.
The five squares are the five credits. The mark is monochrome and carries no colour dependency, so it survives one-colour printing, foil, blind deboss and a 12-pixel retail thumbnail. It is drawn as vector text so it can be set in any size without a raster asset.
It is not a certification, and no body audits it. It asserts one thing only: a complete five-line credit is published in this book. A book whose Models line names four models is exactly as entitled to the label as one that says none.
That is the design. A mark that meant "no AI here" would be a purity claim, would need an auditor, and would be gamed within a season. A mark that means "the credits are on page four" can be checked by opening the book.
FiveFor publishers: how this maps to ONIX
Most of this standard can already be filed in ONIX today. Two things cannot, and they are the two that matter.
On 10 August 2026 EDItEUR published an application note, Aspects of AI in ONIX, describing how AI involvement should be expressed in book records. It is sound work and this proposal is written to sit on top of it, not beside it. Its disclosure threshold — disclose anything beyond de minimis involvement — is adopted here verbatim.
| Credit | Expressible in ONIX today | Status |
|---|---|---|
| Concept | A09 Created by, A38 Original author |
Approximate. Usable. |
| Direction | Nothing. Z01 "assisted by" attaches a tool to a human, which states the
opposite relationship. |
Gap — new role code needed |
| Models | <UnnamedPersons>09</UnnamedPersons> — "generative AI",
anonymous by construction |
Gap — the model cannot be named |
| Editing | B01 Edited by | Covered |
| Art direction | A36 Cover design or artwork by, A11 Designed by,
A12 Illustrated by | Covered |
The inconsistency this proposal exists to fix
ONIX can already name a synthesized audiobook narrator. A generic synthetic voice may carry a
brand name through <AlternativeName>. If that voice was trained on a specific
human performer — an authorized voice replica — the standard can name
the performer.
For the text of a book, the only available value means an unnamed generative AI.
Audiobooks get actor credit. Prose does not. That is not a principle anyone argued for; it is an artefact of the order the codes were written in. The machinery for naming a non-human contributor already exists in the standard — the ask is to permit it for text.
Two proposed additions
- A contributor role for AI direction. A new code in ONIX Code List 17 for the
human who selected and steered generative models — provisionally "AI direction
by". This is a human contributor role, distinct from
Z01"assisted by", and it is what makes accountability legible. - Permission to name the model. Allow a generative model to be identified by
name and date on text roles, using the same
<AlternativeName>mechanism ONIX already permits for synthetic voices, with the vendor in<CorporateName>. The alternative — anonymity by design — is the only disclosure regime in publishing that deliberately withholds the name of a contributor.
Code list values shown here are placeholders. Assigning them is EDItEUR's to do, not ours. Comments and corrections on this section are actively wanted — particularly from anyone who files ONIX for a living.
SixThis has been done twice before
Film, 1933–1941 — and it was not organic
Early films credited nobody; studios withheld actor names to hold salaries down. When credit did arrive it was unreliable: it was sometimes handed to a producer's favourite or to a writer who had never touched the script, and working writers often learned what credit they had received by watching the titles at a preview screening.
In April 1933, one hundred and seventy-three screenwriters founded the Screen Writers Guild, with control of credit as a founding purpose. Their first collective bargaining agreement, in 1941, formalized credit control and mandatory arbitration. The remedy was never a purity test. Nobody proposed certifying a film as wholly written by one man. The remedy was a list of who did what, adjudicated by someone other than the party paying the bills.
It worked so completely that we forgot it was contested. No moviegoer believes the lead actor wrote the screenplay. Nobody experiences a credits roll as a confession.
Science, 2012–2022 — and it needed no union
Academic publishing had the same problem in a text medium: a single "author" label that had
stopped describing who did what. The fix was CRediT, the Contributor Roles
Taxonomy — fourteen roles, begun at a Harvard and Wellcome Trust workshop in May 2012,
published in 2014, and ratified as ANSI/NISO Z39.104-2022. It is now used by more
than fifty organizations across thousands of journals.
Authors have no guild that can halt a publisher's production line, so the film route is closed to us. CRediT shows the open one: publish the taxonomy, let it be adopted voluntarily, and let a standards body ratify what practice has already settled.
SevenObjections
The serious ones, including the one we cannot answer.
It is self-reported. Liars will lie.
Yes. This is the real weakness and it should be stated plainly rather than buried. Film credits are arbitrated by a guild; CRediT is gated by journals; a book credits page has neither backstop, and the authors flooding the market with generated books are the least likely to fill it in.
The answer is not that itemization is enforceable. It is that itemization is a better norm than certification even unenforced, because it gives an honest author something to say instead of something to swear. Attestation asks you to prove a negative about a line nobody can define. A credit asks you to describe what you did. One of those a decent person can complete accurately; the other invites hairsplitting about spell-checkers.
There is also a hook already in the law. In the United States copyright requires human authorship, so a false Direction credit is not merely embarrassing.
Models are versioned, deprecated and silently updated. The credit rots.
So do actors. Christopher Plummer was thirty-five when The Sound of Music opened and eighty-eight when he stepped in for Kevin Spacey in All the Money in the World — physically and vocally two different instruments, credited under the same name both times, and nobody was misled.
A credit is pinned to one release with one date. That is why this standard requires the model family and the month rather than a build number. Name the performance, not the career, and let the name age honestly.
Will disclosure kill sales?
Unknown, and nobody should pretend otherwise. Current sentiment runs against generated prose while softening on research and outlining, and that may move either way.
But the question is misaddressed. Whether a reader wants to spend money on a book a machine helped write is a market decision. It has never been the publisher's decision to make on the reader's behalf, and the present arrangement — where a retailer requires disclosure at upload and the reader is told nothing — makes it one.
Isn't this just a scarlet letter for AI users?
It is the opposite, and that is why the label reads credits published rather than human made. Under a certification regime, the only available public statement is an assertion of purity, so every author who used a model has an incentive to stay silent. Under an itemization regime, disclosure is the ordinary case and there is a normal way to be honest.
Five lines is also less than a film's credit roll, less than a journal article's contributor statement, and about the same as the cataloguing data already printed in most books.
Why not just wait for EDItEUR, NISO or the Authors Guild?
EDItEUR has already moved, and this proposal adopts its threshold and defers to it on code assignment. But metadata is filed by publishers and read by machines; the reader never sees it. A credits page is the reader-facing half, and it needs no committee to start.
CRediT was in use for eight years before it was ratified. Practice first.
Who is behind this?
One author, with a stake to declare. See this page's own credits, which are filed under the standard being proposed — including the machine-generated artwork at the top of the page.
The proposal is offered for adoption, revision or replacement by anyone. It is not owned in any way that matters, and it would be better if a body with standing took it over.
EightAdopt it
If you are a self-published author
You can do this today, without anyone's permission, and it costs you nothing but honesty. Put the five-line block on your copyright page. Fill in none on the Models line if that is the truth — it is worth more than a badge, because you wrote it. Use the label if you want it. Where your retailer asks you to declare AI use at upload, declare it there too; those answers go to the retailer, not to your reader, and only your credits page reaches the person who bought the book.
If you are a publisher
Add the block to your copyright-page template and ask for it in your author questionnaire, in
place of — or alongside — an AI warranty clause your authors cannot reliably interpret.
File what you can in ONIX now: contributor roles for concept, editing and art direction, and the
existing <UnnamedPersons>09</UnnamedPersons> value where a model was used.
If you file ONIX at scale, the two proposed additions in section Five are worth your comment.
If you are a standards body
Take it. Change it. Rename it. Assign real code values and drop the placeholders. The five roles matter; the authorship of the list does not. The most useful outcome of this page is that it becomes redundant.
If you are a retailer
You already collect AI disclosure at upload. Publish it. A field on the product page, drawn from metadata that already exists, would do more for reader trust in a week than the detection arms race has managed in a year.
NineThis page's credits
Filed under the standard it proposes. It would be difficult to argue for disclosure and then not disclose.
- Concept
- Joe Kuefler
- Direction
- Joe Kuefler
- Models
- Claude Opus 5 (Aug 2026) — research, drafting and revision of this page from the
author's outline, argument and edits.
GPT Image 2 (Aug 2026) — the illustration at the head of this page, generated from a written prompt. It depicts nothing real. - Editing
- Joe Kuefler
- Art direction
- Joe Kuefler — layout, typography and the seal are drawn, not generated
The five-line block above took about ninety seconds to write and answers, for this page, every question the industry has spent a year failing to settle by other means. That is the entire argument.