Is Your Writing Being Used to Train AI?
Galley does NOT use your writing to train AI. For other tools, the answer is in the licence grant in your tool's terms, not on its marketing page. Look for what rights you granted, whether there is an improve our services clause, and who the sub-processors are. Be sceptical of any hosted app claiming your text never leaves the server, because it has to read your text to do its job. What you can reasonably demand is a written promise, a short sub-processor list, and a working export.
This has quietly become one of the most searched questions about writing software, and the honest version of the answer is not the one anybody’s landing page gives you. It is not “yes” or “no”. It is “read the licence grant”, and almost nobody does.
Here is how to find out for whatever tool you are using, including mine.
Find the licence grant
Open the terms of service and search for the word “licence” or “license”. You are looking for a sentence in which you grant the company rights over your content. Something in this shape is normal and necessary:
You grant us a licence to host, store, reproduce and display your content for the purpose of operating the service.
That is a tool saying it needs permission to put your words in a database and send them back to your screen. Without it, it cannot function.
What you are watching for is scope creep in that same sentence. Words like “modify”, “create derivative works”, “sublicense”, “worldwide, perpetual and irrevocable”, or “in any media now known or later developed” go far beyond running a text editor. Sometimes the drafting is lazy rather than sinister. It is still what you agreed to.
Find the improve-our-services clause
The second thing to search for is “improve”. A clause allowing a company to use your content “to improve our services” is the one doing the work in most privacy policies written before 2022, because at the time it meant crash reports and usage patterns. It now comfortably covers training a model, and nothing in the sentence has to change for that to be true.
If a tool has added AI features since you signed up, that clause is where the permission came from.
Find the sub-processor list
Under GDPR, a company processing your data has to name the other companies it passes it to. That list is usually a page called “sub-processors” linked from the privacy policy. Read it.
You are not looking for anything scandalous. You are counting. A short list means few places your manuscript exists. A long list with three model providers on it tells you where your prose is going, regardless of what the marketing page says about your work being yours.
If a tool has no such list, that is its own answer.
Be sceptical of the strongest claims
A hosted writing app cannot honestly promise that nobody could ever read your text, and you should treat it as a red flag when one does.
Mine cannot either, and I’ll be blunt about why. Galley reads your manuscript to build your PDF, lay out your EPUB, and run the name-consistency check. That is server-side work on plain text. It is not end-to-end encrypted, it cannot be while it does those things, and any tool that formats your book and claims otherwise is either wrong or using “encrypted” loosely.
What protects you is not cryptography. It is a promise, a short list of who else touches the data, and the ability to leave with everything.
What I do, so you can hold me to it
Nothing in Galley writes, rewrites, suggests or completes a sentence. Nothing in it trains on your book, not to improve a model and not to improve Galley. There is no model provider in the stack to send your prose to, which is a stronger guarantee than a policy, because a policy can be revised in a changelog and an architecture cannot.
The tools that do exist count, check and typeset: overused words, name consistency, hyphenation, page counts. None of them generate.
And the part that makes the rest credible: you can export everything at any time, in open formats, including the whole book and its story bible as one zip. Leaving being easy is what makes staying a choice. The whole position, in one page: hard to lose, easy to leave.
Credit where it is due
I am not the only one who has taken a position here, and pretending otherwise would be exactly the kind of marketing this post is arguing against. Novlr is writer-owned and has been publicly anti-AI for years. The Novel Factory is AI-free and has deeper plotting tools than I do.
Two more are worth understanding rather than dismissing. Novelium shares the rule that AI must not write your prose, and then uses AI to read your manuscript and check it, which is a different line from mine but an honest one, clearly stated. Sudowrite is openly an AI that writes prose with you. It tells you so on the homepage. A tool being upfront about what it does is not the problem this post is about.
The ones to worry about are the tools that do not say.
The short version
Do not ask a company whether it trains on your writing. Ask what its terms permit it to do, count its sub-processors, and check you can get your manuscript out. The answer to the first question can change with a product update. The other two are much harder to quietly reverse.
The no-AI position · Why Galley will never put AI in your prose · Philosophy
More on galley
- Galley for Self-PublishingWhat self-publishing a novel in Galley involves: a print-ready interior PDF, a valid EPUB, a tuned file per store, and two checks that run before you upload to KDP.
- Galley for Traditional PublishingQuerying agents or working with an editor: standard manuscript format, the submission package, watermarked beta drafts, and getting an editor's Word comments back where they belong.
- Why Galley Will Never Put AI in Your ProseAI writing tools are everywhere. Galley deliberately has none in your manuscript: no generation, no rewriting, no training on your words. Here's the reasoning, in full.