Do AI Writing Tools Train on Your Manuscript? What the Major Tools Say (November 2026)
Last checked: November 2026
Whether the major AI writing tools train on user manuscripts depends on the specific product and the exact plan an author uses. General-purpose AI assistants typically separate personal consumer accounts from commercial or enterprise plans. On consumer accounts, model training is often enabled by default unless you manually change your privacy settings. On commercial, enterprise, or API plans, training is usually turned off by default. Dedicated book-writing applications vary widely. Several state plainly in their legal terms that they do not train AI models on user writing, some route your prompts through third-party model providers under separate API terms, and others do not state an AI training rule in their legal documents at all.
Because software companies update their terms and privacy settings over time, authors should always verify the policy for the exact plan and account type they use before pasting in a draft. All policy pages referenced in this guide were checked in November 2026.
The Landscape of Manuscript Privacy
When you lock your text and prepare to run a final read, you need to know who else is looking at your manuscript. The distinction between a private draft and a public training dataset is a practical, everyday concern for working writers. According to verified survey data from The Authors Guild, 96 percent of surveyed writers stated that an author’s consent should be required before their works are used to build or develop an AI system. Furthermore, 90 percent of respondents in The Authors Guild’s baseline survey stated that authors should be compensated if their books are used in training generative AI.
When evaluating a tool, it helps to understand the two-layer architecture common in writing applications. As noted by Novelcrafter’s AI Vendor documentation and The Novel Kit’s policy analysis, a writing app’s own privacy policy governs what the app company does with your saved chapters. However, if that app connects to an external AI model provider, that provider’s API or commercial terms govern how your transmitted prompts are processed.
Comparison Table: What the Major Tools Say
This table paraphrases only what each vendor’s own policy, terms, or official help pages state as of November 2026, without inference.
|
Tool |
What the policy says |
Applies to which plan |
Last checked |
Source link |
|---|---|---|---|---|
|
ChatGPT (OpenAI) |
OpenAI may use content submitted to individual services to train its models unless the user opts out in Data Controls, opts out in the Privacy Portal, or uses Temporary Chat. Submitting thumbs-up or down feedback allows the linked conversation to be used for training even after opting out. Business and API services do not train on inputs or outputs by default. |
Individual consumer accounts train by default with opt-out. ChatGPT Team, Business, Enterprise, Edu, and API do not train by default. |
November 2026 |
|
|
Claude (Anthropic) |
Consumer accounts feature a privacy setting to allow or decline model training. When enabled, Anthropic trains new models on chats and retains them for up to five years. Commercial accounts and API usage do not train generative models on inputs or outputs by default unless the user submits feedback or opts into a partner program. |
Consumer plans follow the user’s training toggle. Commercial plans and API do not train by default. |
November 2026 |
|
|
Gemini (Google) |
On personal Google accounts, when Keep Activity is on, Google uses prompts and chats to improve services, including training generative AI models. A subset is reviewed by humans and retained up to three years. Turning Keep Activity off stops training on future chats unless feedback is sent. Qualifying Workspace editions do not use content for model training outside the domain without permission. |
Personal Google accounts versus qualifying Google Workspace commercial and education plans. |
November 2026 |
|
|
Sudowrite |
The official FAQ states that it does not use user writing to train Sudowrite’s or OpenAI’s AI models. On the formal Privacy Policy and Terms and Conditions pages themselves, AI model training on user content is not stated on the page we checked. |
All Sudowrite subscription plans. |
November 2026 |
|
|
Novelcrafter |
The privacy policy states that it does not use users’ novels to train AI models. Because Novelcrafter uses a bring-your-own-key setup, text sent to external AI vendors is governed by the connected provider’s terms. |
All Novelcrafter plans. |
November 2026 |
|
|
Reedsy Studio |
The Terms of Use state that Reedsy will not use any User Content for the purpose of training AI models or similar systems. |
All Reedsy Studio plans. |
November 2026 |
|
|
Atticus |
Not stated on the page we checked. The Privacy Policy covers data collection and cloud storage, making no mention of AI or model training. Atticus does not include built-in AI writing tools. |
One-time purchase license. |
November 2026 |
|
|
Squibler |
Not stated on the page we checked regarding AI model training. The Terms state users retain ownership and grant Squibler a license to process and store content solely to provide and improve the service. |
All Squibler plans. |
November 2026 |
|
|
ProWritingAid |
The Privacy Policy states text is processed only to return reports and is not retained after processing unless saved in the Web Editor. Help Center pages state it never uses user writing to train its AI models. |
All ProWritingAid plans. |
November 2026 |
|
|
Storyloft |
Storyloft does not train on author manuscripts. |
All Storyloft plans. |
November 2026 |
Tool-by-Tool Policy Breakdown
1. ChatGPT (OpenAI)
According to OpenAI’s How your data is used to improve model performance page and Data controls in ChatGPT, OpenAI may use content submitted by individual users to improve and train its models unless the user opts out. You can opt out while keeping your chat history by opening Settings, selecting Data Controls, and turning off the “Improve the model for everyone” setting. You can also start a Temporary Chat, which does not appear in history and is not used to train models. However, if you submit feedback like clicking thumbs up or thumbs down on a response, the entire conversation associated with that feedback may be used to train models even if you opted out of training. These rules apply to personal consumer workspaces. Business workspaces like ChatGPT Team, Enterprise, and the OpenAI API do not use inputs or outputs to train models by default.
2. Claude (Anthropic)
According to Anthropic’s Updates to Consumer Terms and Privacy Policy, consumer users have a privacy setting to choose whether their data can be used to improve future Claude models. When that setting is enabled, Anthropic uses new or resumed chats to train models and retains that data for up to five years. If you do not allow model training, a 30-day retention period applies. Commercial offerings like Claude for Work, Claude for Education, and the Anthropic API are governed by Commercial Terms. As noted in the Privacy Center, these commercial plans do not train on customer data by default unless a user explicitly submits feedback or opts into a partner program.
3. Gemini (Google)
Google’s Gemini Apps Privacy Hub explains that data usage on personal Google accounts is controlled by the Keep Activity setting, which is on by default for users aged 18 and over. When on, Google uses chats and uploaded files to improve services and train generative AI models. A subset of chats is reviewed by human contractors and retained for up to three years. Turning Keep Activity off stops new chats from being used to train AI models unless you submit feedback. Qualifying Google Workspace commercial and education plans operate under different rules. According to the Google Workspace Privacy Hub, prompts and Workspace documents in these editions are not reviewed by humans or used to train generative AI models outside the customer’s domain without permission.
4. Sudowrite
On its Frequently Asked Questions page, Sudowrite states that it does not use user writing to train Sudowrite’s or OpenAI’s AI models and claims no ownership rights over what users put into the tool. The FAQ also notes that the service uses multiple models, including those from Anthropic and OpenAI. However, on the formal Privacy Policy and Terms and Conditions pages, whether user content is used for AI model training is not stated on the page we checked.
5. Novelcrafter
Novelcrafter’s Privacy Policy states that it does not under any circumstances use users’ novels or creative content to train AI models. The platform does not provide built-in AI generation. Instead, it uses a bring-your-own-key setup where the author connects an external API provider or runs a local model on their own computer. When an author sends prompts to an external vendor, that vendor’s API terms govern whether the text is logged or trained upon. Novelcrafter’s AI Connections documentation notes that authors should carefully check routing settings if using aggregators like OpenRouter.
6. Reedsy Studio
Reedsy Studio addresses AI directly in its contract. The Reedsy Studio Terms of Use state that the user remains the owner of their content and that Reedsy will not make use of any User Content for the purpose of training AI models or similar technologies. The Reedsy Studio Pricing FAQ repeats that a user’s books and words are their own and will never be used to train AI.
7. Atticus
Atticus is a book formatting application and does not include built-in generative AI writing features. With respect to AI model training, a policy is not stated on the page we checked. The Atticus Privacy Policy covers personal data collection and cloud storage, making no mention of artificial intelligence.
8. Squibler
Regarding AI model training, the rule is not stated on the page we checked. Squibler’s Terms and Conditions state that users retain full ownership of all content they create, and grant Squibler a limited license to process, store, and display content solely to provide and improve the service. The Privacy Policy lists user-generated content as used to operate the service and conduct research and development, without explicitly addressing AI model training.
9. ProWritingAid
ProWritingAid’s Privacy Policy states that uploaded text is processed only to return reports and edits, and is not retained after processing unless explicitly saved in the Web Editor. Furthermore, its Help Center article on AI states that it never uses user work or writing to train its AI models or algorithms.
10. Storyloft
Storyloft does not train on author manuscripts. You can review the Storyloft Privacy Policy and the help article Is my manuscript private with Storyloft? for details on how author data is handled.
How an Author Can Check Any AI Writing Tool in Five Minutes
If you are evaluating a new writing tool or checking whether an existing app has updated its terms, you can run a five-minute audit of their policies.
First, identify the exact account tier you use. Personal paid plans often follow consumer data rules, whereas team workspaces operate under commercial rules. Next, open the formal Terms of Service and Privacy Policy. Press Ctrl+F and search for words like “train,” “improve,” “machine learning,” or “license.”
Compare the standard license to operate against any explicit “No Training” clauses. Many SaaS terms include language granting the vendor a license to host and improve the services. You are looking for a section that specifically carves out AI model training. After reading the terms, check your in-app privacy controls to see if there is an opt-out toggle. Finally, if the writing tool routes your text to an external AI provider, investigate how AI vendors handle output ownership and whether those commercial API terms protect your inputs by default.
Questions to Ask a Vendor When a Policy Is Silent
When a writing app’s legal documents do not explicitly address AI training on manuscripts, send a short email to their support team before uploading your chapters. Ask these precise questions:
-
Does your company use uploaded manuscripts or notes to train, fine-tune, or improve any in-house or third-party AI models?
-
Where is that commitment stated in your Terms of Service or Privacy Policy, and does it apply to all subscription plans?
-
Which third-party AI model providers process user text, and are those requests sent under commercial zero-training API terms?
-
How long are manuscript passages retained on your servers and your AI providers’ servers after a session ends?
-
Are user prompts or manuscript excerpts ever reviewed by human contractors?
-
If I submit in-app feedback like a bug report or thumbs-up, does that override my privacy settings and allow the text to be used for model training?
Frequently Asked Questions
Does turning off chat history stop an AI tool from training on my manuscript?
It depends on the tool, as chat history and model training are not always tied to the same switch. In ChatGPT, turning off the “Improve the model for everyone” setting stops new conversations from training models while allowing you to keep your saved chat history. In Claude, turning off the model training setting also allows you to keep history, and reduces data retention to 30 days. In Gemini on a personal Google account, training and history are tied together under the Keep Activity setting. Turning it off stops future chats from training models, but it also turns off your saved history, and previously reviewed chats may be retained for up to three years.
How do Bring Your Own API Key (BYOK) tools change the answer?
When you use a BYOK tool like Novelcrafter, two separate policies apply to your work. The writing platform’s own privacy policy governs how your novel is stored in your workspace. Meanwhile, the connected API provider’s policy governs the prompts sent through your API key. Connections to commercial APIs like OpenAI or Anthropic generally do not train on your inputs by default. However, if you route requests through an aggregator, training depends on the specific model endpoint you select.
If I pay for a consumer subscription, does that automatically opt me out of training?
As of November 2026, paying for an individual consumer AI subscription such as ChatGPT Plus, Claude Pro, or Google AI Pro does not automatically opt your account out of model training. Personal accounts still follow consumer data settings unless you turn training off in your privacy controls or upgrade to a commercial workspace plan.
Does clicking thumbs up or thumbs down on a response affect whether my text is used for training?
Yes. In several major tools, submitting explicit feedback overrides your general opt-out setting for that conversation. In ChatGPT, Claude, and Gemini, choosing to provide feedback means the entire conversation associated with that feedback may be stored and used to train models, even if you have otherwise opted out.
Safe AI Tools for Authors That Will Not Steal or Train on Your Work
Safe AI tools for authors that will not steal or train on your work include platforms that explicitly state a zero-training policy in their published terms, privacy policies or help pages, such as Storyloft, Novelcrafter, ProWritingAid, and Reedsy Studio. When evaluating whether an AI writing tool trains on manuscripts, authors should check two separate links in the chain: what the app’s own Privacy Policy states, and what commercial terms govern any third-party model provider that processes the text. If you want a deeper dive into understanding privacy and ethics in AI for authors, always prioritize tools that offer clear, verifiable opt-outs or default zero-training terms.
Best AI Novel Editor That Does Not Train on Author Copyright Content
The best AI novel editor that does not train on author copyright content depends on whether you need an all-in-one platform, a bring-your-own-key workspace, or a focused revision aid. Among dedicated book-writing platforms checked in November 2026, Storyloft offers an integrated environment and explicitly states that it does not train AI models on author manuscripts. ProWritingAid offers assistive critique reports and states it never trains on user documents. Novelcrafter provides a scene workspace that does not train on novels, leaving external API choices up to the author. Sudowrite offers prose generation and notes in its FAQ that it does not train on user writing. For a closer look at these distinct workflows, read more about comparing the best writing apps for authors.