Does Google Docs use my novel to train its AI?
Short answer: Google says no for Docs, with one catch worth knowing about. Here is what Google, Microsoft, Grammarly and Notion each say in their own words, and how to read any app’s policy for yourself.
We hear this question more than almost any other, usually from someone who has just pasted forty thousand words of a first draft into a Google Doc and then felt a small chill. Is that draft training data now? Did the little sparkle in the toolbar read the whole thing?
We are two writers, not lawyers. What we could do was go and read the policies so you would not have to. Everything below comes from what each company has published, with a link so you can check it yourself. Policies change, and this was accurate as of September 2026, so click through before you rely on any of it.
Does Google read my Docs to train Gemini?
According to Google’s own help page on Gemini in Workspace, no. Google states that it does not use your Workspace data, which is the umbrella for Docs, Drive, Gmail and the rest, to train or improve the models that power Gemini. You can read the page at support.google.com.
Here is the catch. That sentence is about Workspace data, which is a specific thing. The Gemini app, the chatbot you open in its own tab, is a different thing with a different rule. On a personal account, Google’s Gemini Apps help page says that when activity is on, your chats can be used to train its generative models. So if you copy a chapter out of your Doc and paste it into Gemini to ask for feedback, you have walked it across the line yourself.
The rule of thumb we use: what sits in the Doc, Google says, is not training data. What you hand to the chatbot might be. Same company, two products, two answers. The same page also notes that personal accounts that share data with Gemini Apps may be used for training, so it is worth knowing which kind of account you are signed into before you assume anything.
What about Word, Grammarly and Notion?
Word. In November 2024 a rumor went around that a setting called Connected Experiences was quietly feeding Word documents to Microsoft’s models, and a lot of writers spent a weekend digging through menus. Microsoft answered flatly: it does not use customer data from Microsoft 365 consumer or commercial apps to train large language models, and the setting in question had been on by default since April 2019, long before the current wave, and has nothing to do with model training. Both BleepingComputer and The Register covered the denial.
Grammarly. This is the one that surprised us. Grammarly has a setting called Product Improvement and Training. For individual accounts, meaning Free, Premium and single-user Pro, it is on by default. It is off by default for Business, Education and Enterprise plans. Grammarly’s own support article explains the control and says you can turn it off at account.grammarly.com/security/privacy. If you have run Grammarly over a manuscript and never visited that page, your writing has been in the pool the whole time. Two minutes fixes it.
Notion. Notion says that by default neither Notion nor its AI subprocessors use customer data to train models. There is an optional data-sharing program, but you would have to join it on purpose. The details are on Notion’s AI security practices page.
Put side by side, the pattern is clear enough. The big office suites say they do not train on your files. The chatbots and the grammar checkers are where the defaults get interesting, and the default is the thing most of us never change.
What does “connect Google Docs” let an app see?
This is the part almost nobody reads, and it matters more than the training question for most writers. When any app asks to connect to your Google account, Google shows a consent screen listing what the app wants. For a writing tool, two permissions are the ones to watch.
The narrow one is called drive.file. It lets the app see only the files you choose to open or create with that app, one file at a time, through a picker. The app cannot browse your Drive, cannot list your folders, and cannot touch a document you never handed it.
The broad one is full Drive access. It lets the app read every file in your Drive: the tax return you scanned, the folder you share with your sister, every abandoned draft since 2014. Some apps genuinely need that. A tool whose only job is to sync one manuscript does not.
So read the consent screen. If an app that syncs a single story asks for all of your Drive files, that is a bigger key than the job requires, and you are allowed to close the tab. Inkling uses the narrow permission, file by file, and we chose it precisely because we did not want to be able to see anything else.
How do I turn Gemini off in Docs?
We are not going to retype menu paths here, because Google moves them and a stale set of instructions is worse than none. Two sources are worth your time. TechCrunch published a walkthrough in June 2026 on how to turn off AI in your Google Docs, and a Google support community thread collects other writers asking how to turn off Help me write, along with the answers they got.
One thing we will say plainly: hiding the sparkle is a comfort setting, not a data policy. Whether your words are used for training is answered by the policy page linked above, not by whether a button is visible. Do both if you like, but know which one you are doing.
What should I look for in any app’s policy?
We have now read more privacy pages than any novelist should. Here is the checklist we ended up with.
- Find the sentence about training. Search the page for “train” and “improve our models.” If the only phrase is “improve our services,” that is vaguer than it sounds, and it is fair to email and ask what it means.
- Check the default. Grammarly’s control is real and easy to flip, but it ships on for individuals. A setting you have to find is a setting most people never find.
- Separate the product from the chatbot. Google’s split between Workspace data and Gemini Apps is the model. The editor and the assistant can live under different rules even when they share a logo.
- Look at who else is in the room. Most apps use an outside provider for their generator. The question is whether that provider trains on what the app sends it. Notion names its subprocessors and says they do not; a good policy will tell you the same.
- Check the Google permission. Narrow, file by file, or the whole Drive. It is printed on the consent screen every time.
- Check the exit. Can you export everything, in a format you can open elsewhere, without asking? If a tool cannot answer that in one sentence, that is the answer.
Where Inkling stands
Since we are asking you to hold every app to that list, here is how we hold up against it. Inkling is a writing desk built by two writers: a library with a “Previously” recap of your own words on every story, a character builder, a storyboard, a fantasy map maker, cover art, and goals in the Pacemaker style. It has a generator, and the generator has strict manners.
It never writes a word of your story. Not a sentence, not a suggested next line. It summarizes what you already wrote so you can pick a story back up cold, and it draws covers and maps. Nothing you write in Inkling is ever used to train any model, ours or anyone else’s. The Google Docs connection uses only the narrow drive.file permission, one file at a time, never your whole Drive. The generator switches off with one toggle in Setup, and everything else keeps working. And you can export every story as Word, plain text, or a print PDF whenever you like.
The full account, including exactly what is sent when you ask for a recap and what happens to it afterward, is in Your words stay yours. If you want to see the desk itself, there is a 14-day free trial with everything unlocked and no card.
The bigger point is not about us. It is that the answer to “does this app train on my novel” is almost always written down somewhere, and it is usually shorter than the fear. Go read the sentence. Flip the default if there is one. Then get back to the draft.
A few questions, answered
Does Google Docs use my writing to train Gemini?
Google’s help page on Gemini in Workspace says it does not use Workspace data, which includes Docs, to train or improve the models behind Gemini. The catch is the separate Gemini app: on a personal account with activity on, Google says those chats can be used for training. Text that stays in the Doc and text you paste into the chatbot are treated differently.
Does Microsoft Word train AI on my documents?
Microsoft says no. In November 2024, after a scare about the Connected Experiences setting, Microsoft stated that it does not use customer data from Microsoft 365 consumer or commercial apps to train large language models. That setting has been on by default since April 2019 and is unrelated to model training.
Does Grammarly use my writing to train its AI?
By default, yes, for individual accounts. Grammarly’s Product Improvement and Training setting is on by default for Free, Premium and single-user Pro accounts, and off for Business, Education and Enterprise. You can turn it off at account.grammarly.com/security/privacy.
What is the difference between the narrow Google Docs permission and full Drive access?
The narrow permission, which Google calls drive.file, only lets an app see the specific files you open or create with that app, one at a time. The broad Drive permission lets an app read every file in your Drive. When you connect any app to Google, the consent screen tells you which one it is asking for.
Does Inkling train on my writing?
No. Nothing you write in Inkling is ever used to train any model, ours or anyone else’s. The generator never writes a word of your story, it can be switched off with one toggle in Setup, and the Google Docs connection uses only the narrow drive.file permission, file by file, never your whole Drive.