Does AI Train on Your Writing? Each Tool's Policy, Oct 2026

Sometimes, and it depends on which door your chapter goes through. The consumer chat apps (ChatGPT, Claude, Gemini) can use what you paste to train future models unless you turn a setting off, and each of them names that setting differently. The developer APIs behind them don't train on your text by default, and neither, by its own FAQ, does Sudowrite. So the real question isn't "does AI steal my writing". It's "which app am I typing into, and what did I click at sign-up?"

The table below is each service's own words, checked on 21 October 2026, with the switch to flip where there is one.

Does AI train on your writing? The policies, side by side

Checked October 2026: every row above was checked against the vendor's own page on 21 October 2026. Anthropic changed its consumer terms in 2025, so the dates on those pages matter more than any article, this one included.

ChatGPT: one toggle, and it follows your account

OpenAI's help page is plain about the mechanism. When "Improve the model for everyone" is off, your new conversations aren't used to train OpenAI models; they still sit in your history. Signed in, the choice applies across your devices. Signed out, it's saved only in that browser, which is easy to forget if you draft on a library computer.

Note the word new. Turning it off today doesn't reach back into the chapters you pasted in spring.

The API is a different contract. OpenAI's developer docs say data sent to the API "is not used to train or improve OpenAI models" unless you opt in, and has been since 1 March 2023. Abuse-monitoring logs are kept for up to 30 days. That's the route most writing apps use, which is why the same model can be a training risk in the chat window and not in a tool built on top of it.

Claude: you were asked, so check what you answered

In August 2025 Anthropic changed its consumer terms: Free, Pro and Max users choose whether their chats train future models, and existing users had until 8 October 2025 to pick. Say yes and Anthropic keeps the data for five years; say no and it stays on 30-day retention (Anthropic's announcement).

Two exceptions in the privacy centre article (updated 16 March 2026) matter to a novelist:

• A thumbs up or down stores that conversation as feedback for up to five years and it may be used for training even with Model Improvement off.
• A chat flagged for safety review can be used to improve safety detection and enforcement.

Incognito chats are never used, whatever the setting says. If you write violence or abuse and want it out of the training pool, incognito plus no thumbs is the cautious combination.

The terms change doesn't apply to the API, Claude for Work, Bedrock or Vertex, which run under Anthropic's commercial terms.

Gemini and Google Docs: two different Googles

This is where most of the "does Google Docs train AI on your writing" worry comes from, and the answer splits in two.

The Gemini app on a personal account trains on your chats while Keep Activity is on. Google's privacy hub (updated 19 October 2026) says a subset of chats is read by human reviewers, reviewed chats are kept for up to three years, and even with Keep Activity off, chats are held for 72 hours. Paste chapter nine into the Gemini app and it's in that pipeline.

Gemini inside Docs is covered by a different promise. For users on Google AI plans, Google says Gemini in Docs, Gmail and Drive "doesn't use your content to train or improve Gemini or other generative AI models." Business Workspace editions get the same commitment, without the customer's permission, plus no human review. One catch sits in the same page: if you choose to share Docs data with the Gemini app, the Gemini app's terms apply instead.

Google doesn't publish a training statement for the text of a plain Docs file on a free personal account with no Gemini involved. The privacy-company blogs that rank for this query (Proton, Fileverse) read that silence as a risk; they also sell the alternative. What can be said is narrower: Google's written promise covers Gemini's use of your Docs content, not every byte in your Drive.

Writing apps: Sudowrite, NovelAI, Novelcrafter, OpenRouter

Sudowrite answers it directly in its FAQ: "we do not use your writing to train AI models." The same FAQ says it runs dozens of models, Claude and several OpenAI models among them. Its My Voice feature does build a private model from samples you upload, which is the point of that feature and worth knowing before you hand it a manuscript.

NovelAI is the strongest on storage: its privacy policy (effective 10 September 2026) says stories are end-to-end encrypted with a key derived from your password, and NovelAI "cannot read" what it stores. Text you send for generation is processed in readable form and "is not kept". The policy doesn't make a separate statement about training.

Novelcrafter has you connect your own AI vendor account, OpenRouter or OpenAI for instance, per its help. That makes the vendor's policy the one that counts, so read the rows above for whichever key you plug in. The same holds for any bring-your-own-key writing app: the provider behind the key sets the training rules, not the app. We couldn't find a Novelcrafter page that states its own training position.

OpenRouter says in its privacy policy (last updated 31 August 2026) that it "does not use your Inputs or Outputs for model training", but that some model providers it routes to may. Its logging docs let you set, account-wide or per request, that it should never route to a provider that trains. Free models and paid models have separate switches, so check both. If you came here from what OpenRouter is, this is the one setting to change before you paste a chapter.

What this page can't tell you

Two different questions get asked with the same words.

The one above is about the text you type in today. The other is whether your published books were already in a training set. That's a lawsuit question, not a settings question. The largest answer so far: in July 2026 a court gave final approval to the $1.5 billion Bartz v. Anthropic settlement over books downloaded from Library Genesis and Pirate Library Mirror, about $3,000 per work on its Works List (Authors Guild). No toggle reaches back into that. If you have a published book, the settlement site is where claims are handled.

We also left out services whose pages we couldn't read clearly for consumer writing use. Mistral's help page, for one, covers Vibe and Studio but not Le Chat by name.

Where KudoNovel sits

KudoNovel saves your manuscript to your KudoNovel account. Its built-in AI, KudoAI, sends requests through KudoNovel's server to OpenRouter's zero-data-retention hosts that don't train on prompts. There's no key to set up and no routing switch to find: the no-training route is the only one.

Your book is stored on your account behind row-level access rules, saves automatically and syncs across devices. On KudoAI nothing you write trains a model, because the hosts are zero-data-retention, and there is no consumer chat default to keep an eye on.

What to do before you paste chapter one

Open the settings page of the one tool you actually use and find the switch from the table. Flip it, then check it again after the next terms update email, because that's when defaults move. If you'd rather not think about consumer chat settings at all, draft where requests only go out on a no-training route: start a book in KudoNovel on the free credit you get when you confirm your email.