Private AI for writers: where your manuscript goes when you paste it

The Morse teamPublished October 25, 20269 min read

Private AI for writers means getting help editing or brainstorming a book without the manuscript training someone else's model or piling up on their servers. As of October 25, 2026, most consumer chatbots train on what you paste unless you switch it off, and a manuscript also passes through tools and people you don't control: the grammar checker in your Google Doc, the AI side panel, and your editor's or agent's own chatbot account. The Authors Guild now says no one in publishing should paste your manuscript into a consumer chatbot without your written permission.

We read the Authors Guild's guidance and model contract clauses, Google's and Grammarly's own help pages, the Bartz v. Anthropic settlement website, and the US Copyright Office's report on copyrightability on October 25, 2026. ProWritingAid deserves credit first: it says plainly, "We never use your data or your writing to train our AI models or algorithms" (ProWritingAid).

Does AI train on your manuscript?

On most consumer chatbot plans, yes, unless you turn it off. OpenAI says that on Free, Plus, and Pro, "we may use your content to train our models" (OpenAI, archived October 19, 2026). Anthropic may train on Claude chats "unless you opt out through your account settings" (Anthropic), and in the Gemini app, "If you're 18 or over, Keep Activity is on by default" (Google). We list each app's default and switch in which AI companies train on your data. Grammarly's support page says "Product Improvement and Training is on for individual users by default," and the setting covers content in "an application or website where our product is active," which includes the Google Doc you draft in if the extension runs there (Grammarly).

Where does your manuscript go, by how you hand it over?

The same chapter is treated differently depending on the route it takes. Every cell comes from the company's or the Guild's own page, read on October 25, 2026; OpenAI's pages are archived copies from October 19 and 20, 2026.

Is Gemini in Google Docs safer than pasting into the Gemini app?

Yes, by Google's own terms, for users with Google AI plans. Google's help page says that for them, Gemini in Docs, Gmail, Drive, and the other Workspace apps "uses your content in Google Workspace to provide more useful responses to your prompts and doesn't use your content to train or improve Gemini or other generative AI models" (Google).

The same page draws the line at the Gemini app. If you share Workspace data with Gemini Apps, "or through screen actions (including screenshots)," it "may be used for model training and improvement." So a chapter you ask about in the Docs side panel and the same chapter copied into gemini.google.com fall under different rules. Google's page doesn't say which promise applies to a personal account without a Google AI plan, so we can't tell you.

Does turning off training keep your manuscript out?

Not entirely, and the gap runs through the Authors Guild's own advice. The Guild's model clause tells a publisher to keep a manuscript out of training "such as by opting out of allowing training in user settings," and its April 16, 2026 statement says "All of the common chatbots provide this option in their user settings" (Authors Guild). That setting doesn't cover the thumbs buttons. OpenAI says that if you rate a reply, "the entire conversation associated with that feedback may be used to train our models," even after you opt out (OpenAI, archived October 19, 2026). We list the other apps' feedback exceptions in which AI companies train on your data.

An editor who opted out and then rated a helpful summary of your book may have sent the whole chat to training. The Guild's best practices, dated May 11, 2026, already warn that "it is not fully clear how those settings will be honored over time" (Authors Guild). Opting out also doesn't reach back to chats already used, which we cover in does ChatGPT train on your data.

Can your editor or agent put your manuscript into ChatGPT?

Not without your written permission, according to the Authors Guild. Its April 2026 statement says "Editors, agents, and others in the industry who have access to authors' works should not upload any manuscript to or otherwise prompt consumer-facing chatbots with any author's works without first getting the author's written permission." It says doing so "may constitute a violation of the author's copyright or right of privacy." It also asks publishers to "enforce strict policies with their staffs" and to keep internal AI systems sandboxed (Authors Guild).

The Guild recommends two clauses for publishing agreements:

• No Uploading Author's Personal Information or Manuscript to Commercial AI. The publisher "shall not upload the Work or any of Author's personal information to consumer-facing AI systems for purposes such as generating summaries, assessments, or marketing copy without written permission from the author," and if you grant permission, it must keep the manuscript out of third-party training.
• No Substantive Editing with AI. The publisher "will not use AI to substantially edit a manuscript," with basic spelling and grammar checkers excepted.

If your contract is still open, ask your agent about both, and raise the thumbs buttons too, so that permission to use a chatbot comes with a promise not to rate its replies. This isn't legal advice.

Will ChatGPT steal your writing?

It can train on it, which is the risk OpenAI's own pages describe. OpenAI says it takes "steps to reduce the amount of personal information in training datasets" (OpenAI, archived October 19, 2026), which covers personal details, not your plot or your prose. None of the pages we read says what training does with a pasted chapter, or whether a model could repeat it. Whether it's safe to put your novel in ChatGPT comes down to the settings and the buttons you press.

What did the Anthropic settlement cover?

Published books that Anthropic downloaded from pirate libraries, not anything users typed. Bartz v. Anthropic (N.D. Cal., No. 3:24-cv-5417) alleged that Anthropic infringed copyrights by downloading books from Library Genesis (LibGen) and Pirate Library Mirror (PiLiMi) to train large language models. Anthropic "denies all the allegations" and argues its use was fair use. The settlement FAQ says the case "has not gone to trial," and the court "has not decided in favor of" either side (settlement FAQ).

As of October 25, 2026, the settlement website says:

• The money: a $1.5 billion non-reversionary fund, paid in four installments with the last due by September 27, 2027, divided equally across class works with valid claims.
• The books: Anthropic downloaded about 7 million files, and about 40% were duplicates. Only works with an ISBN or ASIN, registered with the Copyright Office within set time limits, are in the class.
• The files: Anthropic must destroy the books it downloaded from LibGen and PiLiMi, and it "represented that neither the LibGen or PiLiMi datasets, nor any portions of those datasets, were in the training corpus of any of its commercially released large language models."
• The limits: it releases only past conduct, up to August 25, 2025. "No claims based on the output of AI models are released," and it isn't a license to train on any book.
• The dates: final approval came on July 20, 2026, and the first payments are expected "on or before November 15, 2026" (settlement website).

A chapter you paste into Claude today falls under its training setting, not this settlement. We go through that setting in is Claude private.

Can you copyright a book you wrote with AI help?

The Copyright Office says AI help doesn't by itself cost you protection for what you wrote. Its report of January 29, 2025 concludes that "the use of AI tools to assist rather than stand in for human creativity does not affect the availability of copyright protection for the output." It also says copyright doesn't extend to "purely AI-generated material," and that "prompts do not alone provide sufficient control" (US Copyright Office). The Authors Guild tells authors to "disclaim any AI-generated text in your work on your copyright registration application" (Authors Guild).

How Morse keeps your manuscript private

Most routes in the table above leave your manuscript on someone else's servers, under a training setting they wrote, with exceptions they chose. Morse is built to keep your chapters off its servers and out of training.

Morse doesn't train AI on your content, sell your data, or show ads, so there's no switch to find before you paste chapter one. Your chat history, every pasted chapter included, is stored in your browser, encrypted, not on Morse's servers. Unlocking it takes two keys, one that never leaves your device and one Morse's server gives you when you sign in. Attached files are sent with the message and never stored by Morse. Your device holds the only copy, so deleting a chat or clearing your browser removes it for good, and Morse can't recover it. Keep your manuscript backed up somewhere you control. On paid plans, an unfiltered model takes on the dark fiction other apps refuse.

Then pick the mode by how private the draft is:

• E2EE, for the unpublished manuscript no one else should read. On Pro, your browser checks a secure enclave's hardware attestation before anything leaves your device, then encrypts each message so only that enclave can read it. Morse's servers relay text they can't read, and each reply has a lock you can open to see every check that passed. E2EE runs open-weight models like GLM, Kimi, and DeepSeek, with no web search or image generation, and it rests on the enclave hardware and attestation working as designed. The enclaves are run by third-party providers, not by Morse. Here's how end-to-end encrypted AI chat works.
• Private, for brainstorming plots, names, and titles with hosts that agreed to keep nothing. Private is the default for a new chat. Open-weight models like DeepSeek, Kimi, and GLM, plus Grok, run with zero data retention, because Morse has zero data retention agreements with its inference, search, and other providers. Where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't.
• Anonymous, for a line edit from GPT, Claude, or Gemini without the provider knowing it's you. Morse's proxy removes your account, email, and IP address and sends the message under Morse's own account. The provider still sees the text and may keep it for a limited time to check for abuse, under its own policy, so send the chapters you'd rather no company hold through E2EE.

In every mode, because we don't store chats, we usually have very little to give authorities. The privacy policy has the details.

Keep a draft private, route by route

• Chatbots: turn training off before you paste, and don't rate replies in a chat that holds your chapters. On Morse, there's nothing to turn off.
• Uploads: delete ChatGPT Library files separately from the chat. Morse never stores attached files.
• Google Docs: ask in the Docs side panel rather than copying chapters into the Gemini app, and check Grammarly's "Product Improvement and Training" toggle.
• People in publishing: ask for the Authors Guild's no-uploading clause, plus a promise not to rate chatbot replies about your book.

Try a chat from the Morse home page without an account, or start free with no card. End-to-end encrypted chats come with Pro, at $19 a month.