Is It Safe to Upload Confidential Documents to ChatGPT?

Whether it’s safe to upload a confidential document to ChatGPT depends on two things: your plan and what’s in the file. The 10-second check: on a personal plan (Free, Go, Plus, Pro), an uploaded file can be used to train future models unless you’ve turned off “Improve the model for everyone,” so a truly confidential document (a client contract, a report under NDA, anything with other people’s personal data) shouldn’t go in as-is. On a Business, Enterprise or Edu account, uploads are excluded from training by default, which is the right home for confidential work. The safe move on any plan is to strip the identifiers first or use a Temporary Chat. Here’s the detail behind that rule.

What happens to a file you upload

This is in OpenAI’s own documentation, not guesswork. The File Uploads FAQ doesn’t answer it on its own: it says whether uploaded files are used for training “depends on the service you are using,” and sends you to OpenAI’s page on how your data is used to improve model performance. That page is the one that draws the line: business services are excluded from training by default, consumer plans are not, so on a personal plan an uploaded file can be used to improve models while “Improve the model for everyone” is on. Switching it off in Settings → Data controls is a clean opt-out for future uploads.

Retention is separate from training. OpenAI’s retention policy says an uploaded file is kept as long as the chat that holds it; when you delete that chat (or your account, or the custom GPT it lives in), the file is removed from OpenAI’s systems within about 30 days, unless it must be kept for security or legal reasons. Two places keep files longer than the chat: Library, where ChatGPT can save a file separately from the conversation (deleting the chat doesn’t touch that copy, so delete it from Library as well), and projects, where a file stays until you delete it or the project. So data retention has a ceiling, but deleting a file later doesn’t undo any training that already happened from it.

The tier difference that actually matters

For confidential data, the plan is the deciding factor. OpenAI’s enterprise privacy page states that content from its business offerings (ChatGPT Business, Enterprise, Edu and the API Platform, among others) is not used to train models by default. That’s the meaningful line: a personal plan treats your upload as potential training material unless you opt out; a business tier doesn’t, by design.

If your organisation handles regulated data, this is also where a data processing agreement comes in: the contract that sets out how a provider processes personal data on your behalf under the GDPR. Business and Enterprise plans can offer one; personal plans don’t. If you’re uploading anything that belongs to clients or patients, that contract is the difference between a sanctioned tool and a quiet compliance problem.

A quick check by document type

• Your own draft, no third-party data (a CV, a cover letter, your notes): fine on any plan, low stakes.
• A contract, financial report, or anything under NDA: business tier, or redact the confidential clauses and figures first. Not a personal plan as-is.
• A document full of other people’s details (a client list, a tenant complaint, medical notes): that’s their personal data, not yours. Anonymise before upload, or don’t upload.
• A scanned form with ID or card numbers: remove the numbers. “My reference number” works as well as the real digits for getting help. Photographing the form instead of uploading the file changes nothing about this, as the rules for photos work out the same way.

Two tactics that make it safe

First, use a Temporary Chat for anything sensitive: it isn’t used for training and doesn’t stay in your history (OpenAI keeps a copy up to 30 days for safety only). Find it via the model/chat menu; menus move, so if the label has changed, look for “Temporary” and tell us. A temporary chat is the simplest one-off safeguard.

Second, redact before you upload. Replace names, account numbers, addresses and reference codes with placeholders. The content you actually need help with is almost never the part that creates the risk. Knowing which categories to strip makes that a ten-second habit rather than a judgement call each time.

Check it before it leaves your machine

Reading a retention policy at the moment you’re about to drag in a PDF isn’t realistic. That’s what the paste checker is for: paste the text of a document and it flags names, addresses, card numbers and IBANs, API keys and lines like “password: …” before anything is sent. The whole scan happens in the page you already have open, and nothing about the document is uploaded to us, so checking it doesn’t quietly create a second copy somewhere. Thirty seconds of checking beats explaining a leaked client file later.

Keep reading

GuidesDoes Claude Train on Your Chats? One Switch DecidesOn Free, Pro and Max, yes if 'Help improve our AI models' is on, and that stretches retention from 30 days to up to 5 years. How to switch it off.GuidesCan IT See Your ChatGPT on a Work Laptop? Even When You're Logged OutOn a company laptop, IT can see ChatGPT use even when you're logged out, on a personal account or in incognito, sometimes uploads too. What each layer sees.GuidesCan Your Employer See Your ChatGPT? What's Visible and What Isn'tPersonal account on your own device and network: no. Company workspace: admins may be able to read and export your chats. A work laptop can be monitored anyway.