Aller au contenu principal
General

AI and Submitted Manuscripts: The Real Problem Is Who Holds the File

The Writers' Union of Canada wants consent before publishers run a submitted manuscript through AI. The real issue isn't the machine reading it, it's where your file goes.

about 2 hours ago8 min read

For two years we have been having the same argument: is a writer allowed to use AI to write? The question is worn through. But while we squabbled over it, the debate quietly shifted to far less comfortable ground. This time it is not the writer holding the tool. It is the publisher. And the text being fed to the machine is not theirs: it is your manuscript, unpublished, sent in on submission, often without a single contract signed.

In early June 2026, The Writers' Union of Canada (the body that represents Canadian authors) relayed a simple demand in its industry bulletin: that publishers obtain prior consent before running a submitted manuscript through an AI system. Automated summaries, reader's reports, slush-pile triage. The alert was picked up by the French trade outlet Actualitté. It builds on a position from the Authors Guild (the U.S. authors' organisation), which warns that feeding a copyrighted work, or an author's personal data, into an AI system without permission may infringe copyright and privacy.

This shift is worth pausing on, because it reshuffles the deck. Every AI dispute so far has targeted books already published, hoovered into training corpora. Here we are talking about something else: the unpublished manuscript, at the exact moment it is most exposed, sitting between your hands and a house that has promised you nothing yet.

The wrong question, one more time

When this kind of news lands, the reflex is to replay the usual war. On one side, those who say: no AI should ever touch a manuscript, full stop, it is our property. On the other, those who answer: reading committees are drowning in submissions, a triage tool that thins the pile never hurt anyone.

Both camps are arguing about the wrong variable. They talk as if "the publisher uses AI" were a single act, accepted wholesale or refused wholesale. But the use of AI is not the problem. What becomes of the file is.

Let me restate it. A publisher has always been allowed to read you quickly. To hand your typescript to an outside reader. To have an intern write up a synopsis. Nothing new, nothing scandalous. What a publisher has never been allowed to do is hand a copy of your text to a third party that keeps it. And that is precisely what happens when an unpublished manuscript is pasted into a consumer chatbot: many of these systems retain what they are given and fold it into their training data. The private read becomes, silently, an irreversible disclosure.

So the scandal is not that the machine reads you. The scandal is that your file changes hands. The useful test is not "AI or no AI," it is "where does the text go, and who said yes." Exactly the same test as in the entire copyright fight: consent, and the destination of the data.

Why this should annoy both camps

This frame comforts no one, which is a good sign that it is worth something.

It annoys the absolutist anti-AI author. Because the line "zero AI, ever, at any stage" does not survive contact with reality. A contained tool that retains nothing, governed by a consent clause, thinning a pile of three thousand submissions, is not the enemy. And let us be honest: many of the people who demand that no editor go near their text with an AI happily paste their own queries, their back-cover copy, sometimes their chapters, into a chatbot of an evening. The asymmetry "AI for me, not for my publisher" does not hold.

It annoys the hurried publisher just as much. Because "we are overwhelmed" was never a licence to leak. The efficiency argument describes a real problem (committees read too much, underpaid, too fast), but it does not excuse the loss of custody. Pasting an uncontracted manuscript into free ChatGPT is not a productivity choice, it is a break in the chain of confidentiality. The house turns a deposit entrusted to it into a piece of data it gives away. That it happens through carelessness rather than calculation changes nothing about the exposure.

So the right line is not "no AI." It is "no disclosure." Explicit consent, plus a tool that does not keep anything. The rest is a matter of stewardship, not of the morality of art.

The parallel no one dares draw: developers

We wave AI in writing around as if it were unprecedented. It is not. Developers have lived this for two years, and they settled the question long before we did.

A developer pastes code into an assistant every day. No one says he has stopped being a developer. He argues with the tool, refuses most of what it offers, decides line by line. But serious companies drew, very early, exactly the boundary we are talking about here: an internal or contained model, fine; pasting proprietary code into a public chatbot that trains on it, a serious offence, sometimes grounds for dismissal. The distinction was never about "does the AI help." It was about where the file lands.

It is the same governance, word for word, that the book chain needs to import. An author's manuscript is source code. You can let a machine read it if the machine stays in the room. You do not throw it out the window.

And since we are on the subject of abuse, let us name the one looming over publishing. Vibe coding, among developers, describes the person who fires off prompts, accepts whatever comes out without understanding it, and ships. He is the profession's punching bag. There is an editorial equivalent now being born: the editor who lets the AI decide what is worth reading, banks the verdict, and never opens the file. A reading committee that no longer reads. If that is what efficiency means, then the editor's craft is scuttling itself, and no author needs that.

Why this moment, and not another

It helps to understand why the Canadian union insists on the submission stage rather than on published books. Because it is the least protected friction point in the whole chain.

A published book is fenced in: a contract, clauses, sometimes now a dedicated AI clause (The Writers' Union of Canada has in fact added one to its model trade book contract). The manuscript on submission, by contrast, floats in a legal void. It is protected by copyright, of course, but no document states what the publisher may or may not do with it while evaluating it. That is the blind spot. And that is where triage AI slips in, all the more easily since, at the 2026 London Book Fair, some editors already admitted using tools to generate manuscript summaries.

The Canadian union also flags a point we tend to overlook: a licensing market for these AI rights is emerging, and casual use can sabotage it. In other words, giving your manuscript away to a model is not just exposure. It is selling off, for nothing, an economic value that is only beginning to be built for authors as a whole.

What promises of "trust" are actually worth

The houses will tell you to trust them, that no one would do such a thing. The trouble is that the Authors Guild issued its warning precisely after a report that editors were already pasting manuscripts, and authors' personal data, into consumer chatbots. So it is not a theoretical fear. It is an observation.

And since a note of irony is in order: that same guild launched a "human-authored" certification. Except it does not check the text. It takes the author's word for it. An honour-system label, in the middle of a trust crisis whose whole point is that we no longer take each other's word. It tells you how far the sector still is from finding its bearings.

The timeline, for its part, is no longer speculative. The settlement in Bartz v. Anthropic (the U.S. case over using copyrighted books to train a model) runs to around 1.5 billion dollars, for some 500,000 works, roughly 3,000 dollars per book concerned. The fairness hearing was held on 14 May 2026, the claims rate is above 91 percent, and the first payments could arrive in the autumn. AI is not "transforming" publishing in some hazy future. It is already rewriting contracts, settlements, markets.

The only question to ask

For an author, all of this boils down to a very simple test that survives a skim. Before sending a manuscript, or before trusting your pages to any tool at all, do not ask "is there AI in here." Ask: does my text go on to train a model, yes or no?

True for the publisher you submit to. True too for the app you write in. Whether your pages live locally in a Scrivener, or in an online application like WriteControl, Extypis or Dabble, the real dividing line is not the presence or absence of an AI feature. It is the answer to that custody question: do your texts feed the training of anything, or not? A tool can perfectly well offer help on demand and keep nothing. Another can show no AI at all and harvest your data elsewhere. The label tells you nothing; the data policy tells you everything.

The Canadian union is not demanding a ban on AI in publishing. It is demanding two plain, sturdy things: that you be told, and that you be asked. Transparency, and consent. Less spectacular than a manifesto against the machines. Far harder to dodge, though. AI holds the door; it does not cross the threshold unless you let it.

Sources:

HU

Hubert Giorgi

Author

Ready to write with the right tools?

Extypis is a free writing workshop built for authors: structure your projects, manage your characters, and export your manuscripts.

Free, no credit card required