Writing Apps That Don't Train AI on Your Manuscript
Scrivener, iA Writer, Obsidian and Ulysses keep your manuscript on your own device or in a cloud folder you pick. Their makers state that the apps do not send your text to their own servers, apart from optional extras such as a grammar check. Dabble and Ellipsus host your writing in their own cloud but rule out generative AI in writing. Google Docs, Microsoft Word with Copilot and Notion promise not to train on your documents while running AI features beside them. Grammarly is the exception among the big names, because its help center says training on your text “is on for individual users by default”.
I read each company’s own privacy policy and help pages on 13 November 2026. A promise not to train covers only part of the risk. The company that holds the encryption key can still read the file, and Grammarly and Ulysses both describe statistics about your writing that they collect apart from the text.
Training defaults, app by app
| App | Where your text is stored | Training on your content | Source |
|---|---|---|---|
| Scrivener | project files on your computer | none, no AI in the app | Literature & Latte, 29 Jan 2025 |
| iA Writer | your device, or a cloud you choose | none, content processed on device | iA privacy policy, 1 Nov 2026 |
| Obsidian | Markdown files on your device | none, no telemetry | Obsidian privacy |
| Ulysses | your device and your iCloud Drive | “never used to train AI models” | Ulysses help |
| Dabble | your device and Dabble’s cloud | not used to train generative AI | Dabble privacy |
| Ellipsus | Ellipsus cloud | “will never integrate or use generative AI” | Ellipsus |
| Google Docs | Google’s cloud | Workspace content not used for Gemini training | Google Docs help |
| Microsoft Word | your computer or OneDrive | Copilot file contents not used for foundation models | Microsoft Support |
| Notion | Notion’s cloud | not used to train models by default | Notion help |
| Grammarly | Grammarly’s servers | on by default for individual accounts | Grammarly support |
| LanguageTool | LanguageTool’s servers | inputs excluded by default | LanguageTool privacy |
| DeepL Write, free | DeepL’s servers | used “for a limited period” to train | DeepL privacy, section 3 |
| Proton Docs | Proton’s cloud, end-to-end encrypted | Proton cannot read the content | Proton |
Literature & Latte’s knowledge base says Scrivener “contains no artificial intelligence (AI), nor does it participate in any data scraping activity”. It adds that nothing in the background sends your text to another server. The Ulysses privacy policy, last updated 28 September 2026, says your written words, keywords and daily goals are “not saved on our servers”. They are stored as text files on your Mac or iPhone, and in your personal iCloud Drive if you turn on sync. Obsidian’s policy says all data “is never sent to our servers” and that the company collects no telemetry.
Grammarly’s switch is called “Product Improvement and Training”. It covers Free, Premium and single-user Pro accounts, and on accounts bought through Grammarly’s sales team or for education it starts switched off. DeepL’s free translator and free DeepL Write keep submitted texts for a limited time to train its neural networks. DeepL Pro and Write Pro delete texts once the job is done. DeepL’s terms also forbid personal data of any kind in the free versions, which rules out a memoir or a nonfiction book with interviewees’ names. Comparisons of these programs on price and planning features are in Book Writing Software: 12 Programs Compared.
Why authors started asking
In December 2023 the Authors Guild published a survey of more than 2,400 authors in which 96 percent said writers’ consent should be required before their work trains AI. The Guild did not publish how respondents were recruited, so the figure describes its respondents and not all authors.
The Minderoo Centre for Technology and Democracy at the University of Cambridge surveyed 258 published UK novelists and 74 industry professionals between February and May 2025. The university published a summary of the report on 20 November 2025. In it, 59 percent of the novelists said they knew their work had been used to train large language models without permission or payment.
On 20 July 2026 Judge Araceli Martínez-Olguín of the Northern District of California granted final approval to the $1.5 billion settlement in Bartz v. Anthropic, covering 482,460 works at an estimated $3,000 each. That case concerned books downloaded from pirate libraries, and nothing typed into a writing app was part of it.
Who holds the encryption key
A company that encrypts your file with its own key can decrypt it again, whatever its training policy says. Apple’s iCloud data security overview, published 5 January 2026, states that “Standard data protection is the default setting for your account.” Under that default, iCloud Drive is encrypted in transit and on the server with keys held by Apple. Only Advanced Data Protection, an optional setting, makes iCloud Drive end-to-end encrypted with the keys on your trusted devices.
Even with Advanced Data Protection on, Apple’s table lists file type, size and the dates a file was created and last modified among the data that stay under standard encryption. In the United Kingdom, Apple says that users who have not already enabled Advanced Data Protection can no longer do so. Ulysses, Scrivener and iA Writer can all sync through iCloud, and when they do, their privacy depends partly on that setting.
Obsidian Sync, the paid service, is end-to-end encrypted including file names, according to Obsidian’s policy. Proton Docs, part of Proton Drive, is end-to-end encrypted as well, and Proton’s free plan includes 5 GB of storage. Both are options for writers who need a cloud copy and do not want the provider to hold a readable one.
What still leaves your device with training switched off
Grammarly’s help page explains that with training off, your content is not used for training. The same page continues, “We may still collect non-content data, such as writing statistics (total words written, types of suggestions accepted, mistakes made, etc.)”. It adds that this data may still be used for product improvement and training.
The Ulysses policy describes one exception to local storage. With the optional Advanced Check, the checked text goes to Ulysses servers hosted by Hetzner Online and on to LanguageTool, and is deleted shortly after the check. Ulysses also records, pseudonymised, how long and how often texts are checked, along with their language and length.
Google’s help page says that interactions with smart Workspace features are anonymised or aggregated and may be used to improve spell check and autocomplete. Its example is accepting or rejecting a spelling suggestion. A log of when you wrote, for how long and how many words describes your routine in detail, even when no sentence of the manuscript leaves your computer.
Apple Intelligence appears in apps that have no AI
Kirk McElhearn of Literature & Latte wrote on 15 September 2026 that macOS 27 no longer lets you switch off Apple Intelligence as macOS 26 did. Apple’s Writing Tools therefore show up inside Scrivener, although “everything you write in your Scrivener projects stays on your device.” He describes the remaining off switch as Settings, Screen Time, Content & Privacy, Siri, then turning off Writing Assistance.
Ulysses has its own switch, under Edit and Writing Tools on the Mac. Its help page says your texts stay on your device “unless you explicitly choose to send them to Writing Tools”. Apple’s Mac User Guide states that data processed in Private Cloud Compute “is not stored or made accessible to Apple”. The optional ChatGPT extension is different, because Apple’s guide says your request and attachments then go to ChatGPT.
App Store privacy labels are self-reported
Apple’s developer page on privacy details tells developers to describe their app’s practices, “including the practices of third-party partners whose code you integrate”. The same page adds, “You’re responsible for keeping your responses accurate and up to date.” Google’s Play Console help is blunter, stating that “Google’s review process is not designed to verify the accuracy and completeness of your data safety declarations.”
A label marked “Data Not Collected” is therefore the developer’s own claim. The privacy policy is the document a company can be held to, and it usually says more about where text goes than the label does.
If your book is under contract
Your manuscript also lives on your publisher’s computers once you sign. On 16 April 2026 the Authors Guild published a model clause that begins “Publisher shall not upload the Work or any of Author’s personal information to consumer-facing AI systems”. The Guild recommends that publishers get the author’s written permission before any such use.
If you live in the EU
The General Data Protection Regulation gives you a right of access under Article 15 and a right to erasure under Article 17. Article 12(3) sets a response deadline of one month. Article 20 entitles you to receive data you provided “in a structured, commonly used and machine-readable format”, which for a writer means a manuscript export you can open elsewhere. How to rescue a book from an app that has closed is covered in Discontinued Writing Software: How to Get Your Book Out.
Dabble’s policy says it may store your information in the United States. Transfers of EU users’ data to certified US companies can rest on the EU-US Data Privacy Framework, adopted by the European Commission on 10 July 2023. The EU General Court dismissed French MP Philippe Latombe’s challenge on 3 September 2025, and his appeal, case C-703/25 P, was lodged on 31 October 2025. I found no primary source for a ruling on that appeal as of 13 November 2026.
Checking an app before you trust it
Find the privacy policy section on user content and look for where text is stored.
Find out who holds the encryption key, and for iCloud check whether Advanced Data Protection is on.
Search the help center for “training” and note whether the setting is on or off by default.
Read the sections on analytics and crash reports to see what is collected besides your text.
Look for an export to plain text, Markdown, DOCX or RTF before you write the first chapter.
A local app keeps your writing away from the company, which makes your own backups the only protection against a stolen or failed laptop. Mobile options are compared in Writing Apps for iPhone, and the numbers a tracker records are explained in What a Writing Tracker Measures.
Does Google Docs use my novel to train AI?
Google’s help page says it does not use your Workspace data to train or improve the generative models used by Gemini, Search and other systems outside Workspace. Data you choose to share with the separate Gemini apps falls under their own terms and may be used for training.
Does Word's "connected experiences" setting feed my documents to AI?
Microsoft’s documentation, last updated 18 May 2026, says personal data collected from connected experiences “isn’t used to train large language models (LLMs), including those used by Microsoft Copilot.”
Is my manuscript personal data under GDPR?
Invented characters are not personal data. Real, identifiable people are, which includes interviewees in nonfiction and family members in a memoir, and that is why DeepL keeps such texts out of its free version.