Is it safe to upload an unpublished manuscript to ChatGPT?

What happens to what you paste

Three details people often miss:

• Feedback counts. On a personal account, a thumbs up or thumbs down can let OpenAI use that whole conversation for training, even after you have opted out.
• Files can outlive the chat. Uploaded files may be kept separately in your file library; deleting a chat does not always delete them.
• People can access data in some cases. OpenAI lists support, security and legal reasons for authorised access. Opting out of training does not rule this out.

Sources: How your data is used to improve model performance, Chat and file retention, Temporary Chat, Enterprise privacy, API data controls. Checked 20 October 2026.

Claude and Gemini work differently

• Claude (Free, Pro, Max): you choose whether your chats help improve the models, under Settings → Privacy. If you allow it, data can be kept for up to 5 years. If you don't, deleted chats are removed within about 30 days. Team, Enterprise and API use are under separate commercial terms and are not used for training by default. (Anthropic, Claude privacy center)
• Gemini app: with Keep Activity off, future chats are not used for training unless you send feedback. They are still kept for 72 hours, and chats that human reviewers have seen can be kept for up to 3 years. Work and school accounts with enterprise protections are handled differently. (Gemini Apps Privacy Hub)

Training is not the only question

• Patient data. Do not upload identifiable patient data to a personal chatbot account. Using such data with any cloud service needs a legal basis, the right contracts and an approved environment, and your institution decides this, not the chatbot's settings. Removing names and shifting dates helps, but on its own it does not meet formal de-identification standards such as HIPAA's (HHS guidance). Free-text notes, attachments and small subgroups in tables can still point to individuals.
• Your institution's rules. Many universities and hospitals say which AI tools may be used with unpublished data. Check before you upload.
• Shared links. On a personal account, anyone who has a shared link can open that conversation. Check who can see it, and delete links you no longer need. Copies that people have already saved will not disappear when you delete the link (OpenAI).
• Your manuscript vs. someone else's. Authors using AI on their own work is a different question from reviewers uploading a manuscript they were sent in confidence. NIH prohibits its peer reviewers from putting application content into online generative AI tools (NOT-OD-23-149). ICMJE asks reviewers to follow the journal's policy and protect confidentiality (ICMJE, reviewers).
• Disclosure. ICMJE says journals should ask authors to disclose AI use at submission and describe how it was used. AI cannot be listed as an author (ICMJE, authors).
• Patents and confidentiality agreements. If the work contains something you may patent, or is covered by an agreement, ask your technology transfer office before uploading.

Checklist before you upload

1. Check your institution's policy on AI and unpublished data.
2. Use an account where training is off, or Temporary Chat.
3. Remove patient identifiers, and check free text and small subgroups.
4. Send only the sections you need help with.
5. Keep highly sensitive novel details out if you are unsure.
6. Don't use thumbs up or down on chats you want kept out of training.
7. Delete the chat and any uploaded files when you are done.
8. Note what you sent and where, so you can write the AI-use statement.
9. Review every suggested change yourself before it goes into your manuscript.

Will it actually find the problems?

General chatbots are good at language and at obvious inconsistencies. In our own test on a specialist chapter with 18 physician-confirmed errors, the strongest models found roughly 12 to 15. No model found all of them, and some flagged correct text as wrong. Treat the output as a list to check, not a verdict.

Where MediCrab fits

MediCrab is a desktop app for checking medical manuscripts before submission.

• Your files stay on your computer. Manuscripts, data tables, PDFs, check records and edit history are kept in a project folder on your computer, and statistics are computed there.
• You can see what leaves. Content goes to a language model only when you use an AI feature, and every request is listed in the app under "What left this machine".
• De-identification built in. The De-identify tab shows, column by column, what could identify someone, and you can correct any call. The de-identified copy replaces names, phone, ID and record numbers with codes such as ⟦P001⟧ and shifts dates together so intervals stay true. The key stays on your computer and the original file is not changed.
• Models provided through MediCrab. The models that come with a MediCrab account use only providers that commit not to train on your content (listed in section 6 of our Privacy Policy). Providers usually keep requests for up to about 30 days for abuse monitoring. Our relay server does not write manuscript text to its database or logs, and keeps replies in memory for at most 15 minutes so an interrupted connection can resume.
• Your own account. You can also use your own API key, or your own ChatGPT or Claude subscription on your computer. Then that provider's terms for that account apply, and personal subscriptions may allow training unless you turn it off.
• Nothing changes without you. The original manuscript stays read-only. AI edits reach the working copy only after you accept them.

FAQ