Is it safe to upload documents to ChatGPT, Claude or Gemini?

You have a contract, a CV or a support ticket export, and you want an AI assistant to summarize it. The upload button is right there. Before you use it, it helps to know what actually happens to the file, and what you can do so the answer to "is this safe?" does not depend on trust alone.

This is not a scare piece. Modern assistants are useful, and many teams use them responsibly every day. The point is to make the decision deliberate.

What happens to a file after you upload it

When you attach a file to a chat, roughly the same things happen regardless of the provider:

1. The file leaves your device and travels to the provider's servers over an encrypted connection.
2. The server extracts the text (and sometimes images) so the model can read it.
3. The content becomes part of the conversation, which is stored with your account for some period so you can come back to it.
4. Depending on your plan and settings, the conversation may be reviewed by humans for safety, used to improve models, or kept for a fixed retention window before deletion.

Steps 1 to 3 are unavoidable: that is what "using a hosted model" means. Step 4 is where plans and settings differ, and where most of the questions come from.

Consumer plans versus business and enterprise plans

Providers usually separate two worlds:

• Consumer plans (free and personal subscriptions). These are aimed at individuals. Historically, several providers have used conversations from these plans to improve their models unless you opt out, and they offer a toggle for that in the settings. Retention windows and human review policies vary. A 2025 review of six major developers' privacy policies by Stanford HAI found training on consumer chat data to be the default across the board; see Be careful what you tell your AI chatbot and our summary in AI companies train on your chats by default.
• Business, team and enterprise plans, and APIs. These are aimed at organizations. They typically come with contractual commitments: no training on your data by default, defined retention periods, data-processing agreements, and sometimes regional hosting.

The exact wording changes over time and differs between OpenAI, Anthropic and Google, so this article deliberately avoids quoting specific policies. Check the provider's current data controls and terms for the plan you actually use, and re-check when they change. Two questions to answer for your account:

• Is my content used for training by default, and where do I turn that off?
• How long are conversations and uploaded files kept, and can I delete them?

If you cannot answer both, treat the upload as if the answer were "yes" and "indefinitely".

The part that settings do not fix

Even with training turned off and short retention, a few things remain true:

• The document was still transmitted and processed by a third party. For personal data of other people (clients, employees, patients), that is a processing activity you are responsible for, whatever the provider promises.
• Confidentiality clauses do not care about toggles. If a contract says a document must not be shared with third parties, sharing it with an AI service is sharing it with a third party.
• Prompt history is a new copy. Anyone with access to your account, or to a shared workspace, can scroll back and read the file.
• Mistakes are permanent. You cannot un-send a file.

None of this makes uploads wrong. It means the safest upload is one where the sensitive parts were never in the file to begin with.

Anonymize first, then upload

Most documents that people send to an assistant do not need real names, real companies or real identifiers for the assistant to do its job. A contract can be summarized just as well when "Acme GmbH" reads "Company A" and "Jane Doe" reads "Person 1". A support ticket can be classified when the customer's email is email1@example.com.

Replacing those values before the upload changes the risk profile completely:

• The provider never receives the identifying data, so training, retention and human review no longer touch it.
• Confidentiality is easier to defend: what left your device was a de-identified draft.
• Consistent labels ("Person 1" everywhere) keep the meaning intact, so the model still understands who did what.
• A mapping table lets you restore the original names in the answer afterwards.

That is exactly what Unformation does, and it does it without becoming another third party: detection and replacement run inside your browser, and the file never leaves your device. You can verify that with the browser's Network tab or by turning Wi‑Fi off before you drop the file. The security page explains the design in detail.

A checklist before you press send

Use this for every upload, not only the obviously sensitive ones.

Practical rules that hold up over time

Policies move; habits are what protect you. A few that keep working:

1. Default to anonymized input. Make the de-identified version the one you paste, and only fall back to real data with a clear reason.
2. Keep the mapping table local. It is the key to your pseudonyms; treat it like a password file.
3. Prefer organizational plans for work. If your employer offers a business account, use it. The contractual guarantees are worth more than a settings toggle.
4. Turn off training on personal accounts anyway. It costs nothing.
5. Delete conversations that contained sensitive uploads once you no longer need them.
6. Write it down. A one-page team policy ("anonymize before AI, use the company account, do not paste special-category data") removes most of the guesswork.

Short answer

Uploading a document to ChatGPT, Claude or Gemini is safe enough for a lot of everyday work, especially on organizational plans with training turned off. It is safest when the document does not contain identifying data in the first place. Take thirty seconds to replace names, companies and contact details with consistent labels, then upload.

Try it now: anonymize a document in your browser, no upload, no account, works with Wi‑Fi off. Curious how the detection works? Read how it works.