How to Upload a Confidential Document to ChatGPT Safely: the Settings, Step by Step
Published:
On a personal ChatGPT plan, OpenAI says it won’t use an uploaded file to train its models if you switch off “Improve the model for everyone” or upload it in a Temporary Chat (and don’t rate the answer); on Business, Enterprise and Edu workspaces, training is off by default. A setting can’t un-send a file, though, so the step that protects you most happens before the upload: taking out the parts that would hurt if they leaked.
This page is the procedure. If you’re still deciding whether a given file should go into ChatGPT at all, the verdict on uploading documents covers that first. The menu names below come from OpenAI’s help pages as they read on 29 October 2026.
The procedure, in order
1. Cut what ChatGPT doesn’t need. Before you touch a setting, save the document as a new file and work on that copy: replace client names, account numbers, addresses and reference codes with placeholders like “Client A” or “[amount]”. A contract clause reads the same to ChatGPT with “Client A” in it, and the version you upload is no longer the confidential one. If the risky part is a whole section (a price schedule, a medical history), delete it from the copy.
2. Turn off training for your account. On the web: open your account menu, select Settings, then Data controls, select Improve the model for everyone, turn it off and select Done. In the iOS or Android app: open the sidebar, tap your profile icon to open Settings, tap Data controls and turn off Improve the model for everyone. When you’re signed in, the choice follows your account to every device. If you use ChatGPT without an account, the same switch is in Settings, but it’s only remembered in that one browser.
3. Open a Temporary Chat. Open a new chat and select Temporary. Before your first message, ChatGPT asks you to pick Personalized or Unpersonalized. For a confidential file, pick Unpersonalized, which means ChatGPT won’t use what it remembers about you, your custom instructions or your plugins. The choice can’t be changed once the conversation starts. While the chat stays temporary, OpenAI says it isn’t used to train models, stays out of your history and doesn’t create memories.
4. Upload, ask, and leave the thumbs alone. Attach the cleaned copy and ask your question. Don’t rate the answer: OpenAI’s Data controls page says that if you give feedback, such as a thumbs up or down, the entire conversation it belongs to may be used for training, even after you’ve opted out.
5. Close it without saving. A temporary chat can be saved later, and saving turns it into a regular chat that follows your account’s normal settings. Files from a saved temporary chat can also land in your Library. If you just close it, you can’t reopen it from your history, and OpenAI’s Library page says files uploaded in a Temporary Chat are not saved to your account or Library.
What these settings do NOT do
- They don’t make OpenAI forget straight away. OpenAI may keep a copy of a temporary chat for up to 30 days for safety purposes. “Not used for training” and “deleted” are two different promises.
- Turning training off doesn’t delete anything. Your earlier chats, archived chats and project chats all stay where they were. Archiving a chat doesn’t change whether it can be trained on either.
- Deleting a chat doesn’t delete the file. Where Library is available, ChatGPT saves files from regular chats there automatically, and OpenAI states that deleting the chat leaves that saved copy in place.
- Memory is a separate switch. Training and Memory are controlled in different places. With Memory on, and depending on your plan and region, ChatGPT can draw on past chats and on files in your Library to personalise answers. You’ll find it under Settings, then Personalization, then Memory.
- Nothing here checks what’s inside your file. The settings decide what OpenAI does with a document. Whether that document should leave your desk is still your call (and, at work, often your employer’s).
Drafts and personal projects
Your own draft is the easy case. A CV, a cover letter, a chapter of a novel or notes for a talk carry no one else’s data, so the only question is whether you mind the text being used for training. If you do, step 2 covers it, as long as you skip the thumbs up and down; if you don’t, you can skip all of this. The one trap: a draft of a work document (a bid, a board memo, a product plan nobody has announced) belongs to your employer, not to you, so it counts as confidential however rough it is.
Using a work account
On ChatGPT Business, Enterprise and Edu workspaces, OpenAI says it doesn’t use the content to train its models by default. Your organisation’s own rules sit on top: workspace owners can set retention, Memory and access policies, and exporting your data or deleting the account goes through the owner, not through your own settings. Before you upload, check that the account you’re signed in to is the workspace one and not a personal account opened in the same browser. Also check what a workspace admin can see of your chats before a sensitive file goes in.
If you already uploaded it
Work through these in order.
- Delete the chat. In your chat history, open the more options menu (•••) next to the conversation, select Delete and confirm. It vanishes from your account at once, and OpenAI schedules it for permanent deletion within 30 days, unless it was already de-identified and separated from your account, or OpenAI must keep it longer for security or legal reasons. A deleted chat can’t be restored, not even by OpenAI Support.
- Delete the file from Library. On the web, open Library from the left sidebar, select the file and click Delete or the trash icon. If your Library shows Recently deleted, the file waits there and can still be restored; choose Delete forever to remove it now. After that, OpenAI’s systems drop it on the same 30-day schedule.
- Check Memory. Deleting a chat doesn’t necessarily delete a saved memory created from it. Go to Settings, Personalization, Memory and remove anything that came from the document.
- Check projects and custom GPTs. A file added to a project stays until you delete the file or the project; a file uploaded as knowledge to a custom GPT stays until the GPT is deleted.
Deletion removes your copies, on OpenAI’s 30-day schedule. Its help pages don’t say it reverses training that already used the chat, so if “Improve the model for everyone” was on at the time, treat that part as done and switch it off for next time. If the document held other people’s data, tell whoever handles data protection where you work; they’ll know whether it has to go further. Next time, run the file past the list of what never to share before it goes in.
Check the text before you paste it
The Paste Checker catches the parts step 1 is about. Paste in the text of your document and it flags passwords and API keys, card numbers, IBANs and labelled bank account numbers, email addresses, phone numbers, street addresses, several national ID formats and health terms, and it makes a rough guess at people’s names, all inside your browser, so the text never reaches us. It works on pasted text only: it can’t open a PDF or a Word file, so copy the text out first. And it can’t tell that a contract is under NDA. Tick Client or customer data yourself, and it’ll show you how each ChatGPT plan, and the other AI plans we track, treats that kind of file.