For two years we've been replaying the same argument over whether a writer has the right to write with AI, and the question is worn threadbare. Meanwhile, the debate has moved somewhere far less comfortable, since it's now the publisher holding the tool and running your text through it: an unpublished manuscript, sent in as a submission, often without any contract signed.
In early June 2026, The Writers' Union of Canada, which defends Canadian authors, relayed a simple request in its professional newsletter: that publishers get an author's consent before running a submitted manuscript through an AI system, whether to summarize it, write a reader's report or sort the submissions pile. The French outlet Actualitté picked up the alert, which draws on a position taken by the Authors Guild: feeding a copyrighted work or an author's personal data into an AI without permission can infringe copyright and privacy.
This shift deserves attention, because it changes the picture. Until now, all the AI litigation concerned books already published and scraped into training corpora, whereas this is about the unpublished manuscript, at the very moment it's most vulnerable, between your hands and those of a house that hasn't promised you anything yet.
A debate asking the wrong question
When this news breaks, the reflex is to replay the usual war. Some insist no AI should touch a manuscript because it's their property, and others reply that reading committees are drowning in submissions and that a tool to thin the pile never hurt anyone.
Both camps argue as if "the publisher uses AI" were a single act to accept or refuse wholesale, when everything depends on what happens to the file.
A publisher has always had the right to read you quickly, to hand your text to an outside reader or to have an intern write a summary sheet, and none of that has ever caused a scandal. They have never had the right to give a copy of your text to a third party that keeps it. Yet that's what happens when you paste an unpublished manuscript into a consumer assistant, since many of these systems keep what you submit and use it for training, so a private reading quietly becomes an irreversible disclosure.
What causes the problem, then, is that your file changes hands. The useful question concerns where the text goes and who authorized it, and that's exactly the criterion of the whole copyright battle: consent and the destination of the data.
A framework that bothers both camps
This way of looking at things comforts nobody, which is rather a good sign.
It bothers the writer opposed to all AI, whose line, "never any AI, at any stage," doesn't survive scrutiny. A tool that keeps nothing, governed by a consent clause and used to thin a pile of three thousand submissions, can be a useful ally. Many of those who demand that no publisher bring their text near an AI paste their own query letters, blurbs and sometimes chapters into a chatbot in the evening, and demanding AI for yourself while forbidding it to your publisher doesn't hold for long.
It bothers the hurried publisher just as much, because being overwhelmed has never justified letting a text leak. The efficiency argument describes a real problem, that of committees reading too much, too fast and for too little money, but it doesn't excuse the loss of control. Pasting an uncontracted manuscript into the free version of ChatGPT breaks the chain of confidentiality, since the house turns a deposit it was entrusted with into data it hands out, and whether this comes from carelessness rather than calculation changes nothing for the exposed author.
The right line therefore consists in refusing disclosure: explicit consent, and a tool that keeps nothing. The rest is a matter of workflow far more than of morality.
The comparison few dare to make: developers
People talk about AI in writing as if it were an unprecedented situation, whereas developers have been living with it for two years and settled the question long before us.
A developer pastes code into an assistant every day without anyone claiming they've stopped being a developer, since they talk with the tool, reject most of what it suggests and decide on every line. Serious companies still drew, very early on, the boundary at stake here: an internal or controlled model is fine, while pasting confidential code into a public assistant that trains on it is a serious offense, sometimes grounds for dismissal. That boundary has always concerned where the file lands, far more than whether the AI helps.
The book trade can adopt this rule as is. An author's manuscript is worth as much as a company's source code: you can have a machine read it as long as the machine stays in the room, and it must never go out the window.
Since we're talking about abuses, we should name the one threatening publishing. Among developers, the "vibe coder" sends prompts to the machine, accepts what comes out without understanding it and ships it, and the whole profession mocks them. Its equivalent is being born in publishing: the editor who lets AI decide what deserves to be read, accepts its verdict and never opens the file, until the reading committee stops reading. If that's what efficiency looks like, the editor's profession is sabotaging itself, and authors have nothing to gain from it.
Why submission is the most exposed moment
The Canadian union focuses on the submission stage because it's the least protected link in the whole chain.
A published book is governed by a contract, with its clauses, and now sometimes a clause dedicated to AI, like the one The Writers' Union of Canada added to its model contract. A submitted manuscript, by contrast, sits in a legal void: copyright protects it, but no document says what the publisher may do with it while evaluating it. That blind spot is where sorting AI slips in, all the more easily since at the London Book Fair in 2026, publishers were already acknowledging using it to summarize manuscripts.
The union also stresses an often overlooked point: a licensing market is emerging for these uses, and careless practices risk ruining it. Handing your manuscript to a model for free therefore exposes you, and on top of that sells off an economic value that's only just beginning to take shape for all authors.
What appeals to trust are worth
Publishing houses will tell you to trust them and that nobody would do such a thing. The Authors Guild nonetheless issued its warning after a report that publishers were already pasting manuscripts, and authors' personal data, into consumer assistants, so the fear rests on facts.
The same Guild launched a "Human Authored" certification that doesn't check the text and relies only on the author's word. Offering an honor-system label in the middle of a crisis born precisely from the fact that nobody takes anyone's word anymore shows how much the sector is still searching for its bearings.
The timeline, meanwhile, is no longer theoretical. The settlement in Bartz v. Anthropic, the American case over the use of copyrighted books to train a model, covers about $1.5 billion for some 500,000 works, or roughly $3,000 per book. The fairness hearing took place on 14 May 2026, more than 91% of the authors concerned have filed a claim, and the first payments could arrive in the fall. AI is already changing contracts, legal settlements and markets.
The question to ask
For a writer, all of this comes down to a simple test. Before sending a manuscript, or entrusting your pages to any tool, ask yourself whether your text will be used to train a model, far more than whether it will cross paths with an AI along the way.
The question applies to the publisher you're submitting to as much as to the app you write in. Whether your pages stay on your computer in Scrivener or live in an online app like WriteControl, Extypis or Dabble, the dividing line runs through this question of custody: are your texts used, or not, to train anything? A tool can offer on-demand help without keeping anything, and another can show no AI at all while exploiting your data elsewhere, so only the data policy really tells you.
Rather than a ban on AI in publishing, the Canadian union asks for two simple, solid things, that you be told and that your consent be sought, in other words transparency and consent. It's less spectacular than a manifesto against machines, and much harder to get around, since no AI will then have access to your manuscript without your authorization.
Sources:
- Actualitté, "Au Canada, les auteurs veulent un accord avant tout usage de l'IA," 4 June 2026: https://actualitte.com/article/131687/legislation/au-canada-les-auteurs-veulent-un-accord-avant-tout-usage-de-l-ia
- The Writers' Union of Canada, "Industry News, April 2026": https://writersunion.ca/news/industry-news-april-2026
- The Authors Guild, statement on publishers' use of AI (Publishers Weekly): https://www.publishersweekly.com/pw/by-topic/digital/copyright/article/100211-authors-guild-speaks-out-on-publisher-ai-use.html
- The Authors Guild, "Bartz v. Anthropic Settlement: What Authors Need to Know": https://authorsguild.org/advocacy/artificial-intelligence/what-authors-need-to-know-about-the-anthropic-settlement/
- Authors Alliance, "Bartz v. Anthropic Fairness Hearing" (14 May 2026, 91.3% claim rate): https://www.authorsalliance.org/2026/05/14/bartz-v-anthropic-fairness-hearing-observations-and-takeaways/
Hubert Giorgi
Author
The writing studio novelists were missing
Outline, characters, places and objects, the links between them, mind map, scene board, export: everything lives in one place. Plus an AI that knows your story, whenever you want one.
Free to write and to export. No credit card asked.
