Would you hand a stranger your manuscript because he promised he might mention you to his friends someday? That’s the deal most AI chatbots offer writers, and plenty of writers take it.
An author asked in an AI writing group this week which AI she should use to critique her book, and whether the AI would steal her work or use it to train its models. Good question. The answers that piled in were not so good.
Most of the group told her not to worry. AI doesn’t steal, they said. Your book is a drop of water in the ocean. Leave training turned on, they said, so the AI will know you and your style and maybe recommend your book to somebody someday. One commenter said it felt hypocritical to use AI while refusing to let it learn from your data. Another said he’d be proud to have an AI think his novel was good enough to train on.
Oh, hell no.
I turned training off the first day I used AI, long ago, and I’ve turned it off on every AI I’ve used since. If somebody wants me to do work for them, they can pay me. Same with advertising. If a company wants me to carry a pen with its logo on it or wear its T-shirt, it can pay me. I don’t give away free stuff, and a finished manuscript is about the most valuable free stuff a writer could hand over.
The whole debate gets muddled because two separate questions are tangled together. One is whether an AI will copy your book and hand it to someone else. It won’t, in any practical sense. The other is whether the company behind the AI gets to use your book, for free, to make its product better and more valuable. On most consumer chatbots, unless you say no, it does. People lose sleep over the first question. They should be asking the second.
And the stakes aren’t small. AI companies have spent billions of dollars building models out of other people’s writing, and some of that writing was taken without asking. Anthropic agreed to pay authors $1.5 billion in 2025 to settle claims over pirated books it had used. The industry now knows exactly what books are worth. So why would you give yours away?
Will AI steal my book if I upload it for critique?
Not the way most authors picture it. A chatbot doesn’t file your manuscript away and hand chapter seven to the next person who asks. When a model trains on text, it learns patterns across billions of pages. It doesn’t keep a copy of your novel on a shelf.
The group got that part right.
What they skipped is that “it won’t steal it” and “you should let it train on it” are two different claims. If the setting allows it, your chapters help train the next version of the model. Your conversations get kept, and the company’s staff can review them under its policies. None of that is theft in the legal sense. All of it is the company getting value from your work without paying you. Would you let a publisher do that?
How do you stop AI from training on your book?
Find the data or privacy setting and switch it off before you upload anything. Every major chatbot has one, and the names change every few months, so look for wording about improving the model, training, or using your content to develop future models.
On ChatGPT, the setting that lets OpenAI train on your chats is on by default. One commenter in that same thread pointed it out, and she’s right. You have to go find it and turn it off. Claude, Gemini and the others have their own versions of the same switch. Check each one you use, and check again after big updates, because settings screens get rearranged.
If you work inside projects or folders with uploaded files, read what the company says about those too. Some treat uploaded files differently from chat messages. When in doubt, ask the AI itself where the setting is, then go verify it in the settings page.
If an AI company wants to learn from my book, it can pay me for it. I don’t give away free work. – Richard LoweShare on X
Why shouldn’t authors give their books to AI companies for free?
Because your book is your work, and work gets paid for.
I apply that everywhere. A client who wants a book written pays for the writing. A company that wants its logo on my shirt pays for the space. Nobody expects a plumber to fix a leak for exposure, and nobody at OpenAI works for free either. Their engineers get paid. Their servers get paid for. Their investors expect a return. The only people in the whole AI supply chain who are asked to contribute for nothing are the writers whose words make the thing useful. Does that sound fair to you?
The “drop in the ocean” argument doesn’t change that. A single book barely moves a model, true. So does a single dollar barely move a bank. That’s not a reason to hand the bank your dollar.
And the drop-in-the-ocean crowd is also arguing against itself. If your book makes no difference to the model, then keeping it out costs the AI company nothing and costs you nothing. If your book does make a difference, it has value. Get paid for it. Either way, the switch goes off.
Some authors will choose to leave it on, and that’s their call to make. I object to the idea that leaving it on is the generous, sophisticated, obvious choice, and turning it off is paranoid. It’s neither. It’s a business decision about your own property, and the default setting was chosen by the company that benefits from it.
Should authors let AI train on their books so AI will recommend them?
I wouldn’t. The theory in that group was that if the AI knows your book, it’ll recommend you to readers. That’s not how it works. A model that trained on your manuscript learned from the words. It didn’t learn that your book is good, that it’s for sale, or that anyone should buy it.
AI search recommends books based on what’s published about them on the open web: reviews, articles, author pages, interviews, mentions on sites it trusts. That’s a separate process from training, and I’ve explained the difference in training crawls versus live queries. If you want AI to recommend your book, build a public footprint it can find. Don’t feed it your private manuscript and hope.
As for the argument that it’s hypocritical to use AI without giving your data back: I pay for my AI subscriptions. That’s the deal. I pay money, I get a service. Nobody at the software company thinks I owe them my next book on top of the monthly fee, and I’m not going to volunteer it.
What happens to a manuscript after you upload it to an AI?
It depends on the company and the account type. Spend ten minutes reading before you upload a novel.
Business and team plans at the big AI companies keep customer data out of training by contract. Consumer accounts are where the training switches live. Most services keep your chats and files for some period for safety review even with training off, and some let you delete conversations or turn off history entirely. Read the privacy page for your own account type. A blog post about somebody else’s plan won’t tell you.
Treat anything you haven’t published yet with extra care. If a book is under contract with a publisher or covered by a ghostwriting nondisclosure agreement, check that agreement before you upload a word of it anywhere. My clients’ manuscripts never go into a chatbot with training on, and that’s the standard I’d hold any editor or ghostwriter to.
Is it safe to use AI for manuscript critique?
Yes, with training off, and it’s useful. I use AI every day as a digital assistant in my writing, like a virtual assistant only digital, and I still do the writing. Authors in that thread described how they use it for critique, and the methods are sound.
Some set up a project for each book and upload the full manuscript, so the AI can check continuity across chapters. Others ask for an evaluation one chapter at a time and get back a long breakdown of what works and what doesn’t. One author runs continuity checks in NotebookLM and then brings the problems to ChatGPT to fix. Another has Claude check Gemini’s feedback and Gemini check Claude’s. A couple mentioned ProWritingAid and AutoCrit. Both are built for exactly this.
Two cautions. Ask for the hard truth, because chatbots lean toward flattery unless you tell them not to. And keep the final say. AI feedback is a second opinion from a reader who never gets tired and never quite gets it. It’ll catch your missing closing quotes. It won’t tell you whether your book matters. I’ve compared the big assistants in my review of Claude, ChatGPT and Grok.
Who steals self-published books if AI doesn’t?
People do. One author in that thread made the best point of the whole discussion: worry about the humans.
Knockoff artists copy popular books, rewrite them just enough and republish them under a similar name with a similar cover. She said it’s killing nonfiction right now. Pirates put entire books on YouTube, read by AI voices, to collect ad money or steer listeners to shady download sites. And the flood of copycats has made KDP slower and jumpier. Honest authors get caught in the crossfire when Amazon tries to sort out which version is real. That’s part of why I think KDP’s two-book weekly limit punishes the wrong people.
Protect yourself where the theft happens. Search your title and your name on Amazon, YouTube and Google every month. File takedown notices when you find copies; Amazon, YouTube and Google all have copyright complaint forms. Keep dated drafts that prove the book is yours. The Anthropic settlement is worth reading too, because it shows how the courts are starting to treat books taken without permission.
Turn training off today, on every AI you use, before you upload another page. Then use AI for critique all you like. The AI and writing hub covers the rest, and if you’d rather have a human who respects the work read your manuscript, my book coaching does exactly that. Your book took you months or years. A company that wants it can make you an offer.
