Check the service and account before uploading
A privacy decision begins with the exact service receiving the text. Read its terms for your plan and region. Identify training use, retention, human access, sharing defaults, deletion limits and any third-party integrations. If you are using a writing application built on another provider, examine both the application’s policy and the provider relationship it describes. A setting in your personal chatbot account may not govern requests sent by a different application.
OpenAI’s current ChatGPT data-controls guidance says turning off model improvement excludes new conversations from training while retaining ordinary chat history. Temporary chats are not used for model improvement while they remain temporary and may be retained for up to 30 days for safety purposes. [1] Those statements do not amount to a promise of zero storage or a confidentiality agreement tailored to your manuscript.
Anthropic’s commercial-services guidance says chats and coding sessions are not used for training unless the customer participates in specified opt-in programs or provides feedback material; it explicitly distinguishes its consumer products. [2] Verify the policy attached to the product you actually use. Do not apply a commercial-plan assurance to a personal account because the model has the same name.
Send the smallest useful excerpt and remove irrelevant personal information. A continuity question about a train journey rarely requires an entire unpublished novel. If an editor, co-writer or research participant has supplied confidential material, resolve the permission to upload it before doing so.
Keep a recoverable manuscript outside the conversation
Your working manuscript should have a home independent of a chat history. Save ordinary document exports at useful milestones and retain older versions. Synchronization is helpful for moving between devices, but an accidental deletion or corrupt file can also synchronize. Maintain a separate backup whose earlier state you can restore.
A manageable routine is to save a dated working copy after a substantial session, keep a second copy in a separate location and periodically restore a sample file. Check that the restored document opens and contains the expected passage. Include project notes, permissions and cover files in the backup plan, with appropriate access restrictions. A backup that has never been opened is an assumption waiting to be tested.
When you change tools, export before closing the old account. Verify that you have the manuscript and any history you need in readable formats. A new application may preserve the latest text without preserving comments, attribution or earlier revisions. Record what was lost during migration instead of allowing a fresh import timestamp to look like a composition date.
Keep a modest contribution record
A record is most useful when it connects an action to an identifiable version. Separate what the system recorded from what you remember and declare yourself. If you pasted a passage from another tool, identify that origin when known. If the earlier history is missing, label the gap. This is a practical record format, not a legal test or a certification scheme.
| Field | What to record |
|---|---|
| Version and date | A stable filename or identifier, date, and chapter or scene. |
| Starting material | Your prior draft, research source, collaborator’s text or generated output. |
| Assistance | Tool and account type, task requested, and relevant output retained. |
| Your decision | What you accepted, rejected, rewrote or rearranged, and why. |
| Result | The saved version and passages affected. |
| Evidence type | Recorded activity, retained file or author declaration. |
| Limit or gap | Missing earlier draft, uncertain origin or unverified recollection. |
Download the plain-text record template and adapt it to your workflow. For an ordinary revision, a short entry is enough: “Scene 12, version 4. Replaced the generated argument with my own dialogue; retained the arrival paragraph; earlier output saved as draft 3.” Keep the relevant files alongside the entry.
A file hash can help establish that two copies contain the same bytes. It cannot establish who wrote those bytes or whether their expression is original. Timestamps and revision logs also need context. Use them to support an explanation of events, and avoid presenting them as a certificate of ownership.
Plan rights and registration alongside publication
Copyright in qualifying human-authored work arises when it is fixed; registration is not what first creates that protection. [4] Registration can nevertheless be important to enforcement and available remedies. Sections 411 and 412 make timing and other conditions consequential. [5] The copyright guide explains how generated material affects the claim.
Do not substitute mailing a copy to yourself for registration. The Copyright Office explains that this “poor man’s copyright” practice has no provision in U.S. copyright law and is not a substitute. [6] Nor should a provider’s allocation of output rights be treated as a guarantee of copyrightability or exclusivity; OpenAI’s terms expressly qualify their allocation by applicable law and recognize similar outputs. [7]
Keep permissions and licenses for cover art, photographs, substantial quotations and commissioned work with the publication files. Record the permitted formats, territory and duration where the agreement specifies them. A book assembled from individually authorized components still requires attention to the terms of each component.
If you discover a suspicious copy
Preserve the location, date, screenshots and identifying details before contacting the seller or host. Retain your own relevant editions and a comparison of the passages concerned. Distinguish copied expression from a shared premise, ordinary phrase or coincidental resemblance. A measured record will be more useful than a public accusation made before you understand the overlap.
The U.S. notice-and-takedown system allows rightsholders to notify qualifying service providers of claimed infringement and provides a counter-notice process. Notices require specified information and good-faith statements; a dispute can continue after removal or a counter-notice. [8] Use the platform’s appropriate reporting route and seek legal advice where ownership, fair use or the scope of your protected contribution is uncertain.
For a mixed manuscript, be especially precise about the material you claim. A registration or evidence package should not be used to assert ownership over excluded generated text or another person’s material. Preserve the facts, understand the claim and choose a response proportionate to the harm.
Sources & verification
Primary sources checked October 26, 2026. Publication or revision dates below belong to the source; undated pages are identified as such. The craft and workflow recommendations are PennedCanon’s editorial guidance.
- OpenAI, Data Controls in ChatGPTLive help page; model training and temporary chat. Accessed October 26, 2026.
- Anthropic, How do you use personal data in model training?March 16, 2026; commercial offerings only. Accessed October 26, 2026.
- Authors Guild, AI Best Practices for AuthorsMay 11, 2026; professional guidance, not legislation. Accessed October 26, 2026.
- U.S. Copyright Office, Circular 1: Copyright BasicsRevised September 2021; creation, ownership and registration. Accessed October 26, 2026.
- 17 U.S.C. §§ 408–412: registration and remediesCurrent online statutory text; §§ 408, 411 and 412. Accessed October 26, 2026.
- U.S. Copyright Office, Copyright in GeneralUndated FAQ; registration and poor man’s copyright. Accessed October 26, 2026.
- OpenAI, Terms of UseEffective January 1, 2026; Content and Using our services; regional terms differ. Accessed October 26, 2026.
- U.S. Copyright Office, Section 512 notice-and-takedown resourcesLive resource; notice requirements and counter-notices. Accessed October 26, 2026.