Is it safe to upload documents to AI? Check the file before you send it

Uploading a document to AI can save a great deal of reading. Whether it is appropriate depends on the file and the service you are sending it to.

A public report is different from a confidential client agreement. Even an apparently ordinary document may contain comments or earlier edits you did not intend to share.

Before uploading, make a separate copy and check what the assistant actually needs.

1. Check that the file belongs in this service

If it is a work document, use the account and workflow your organisation has approved for that information.

Being able to open the file does not necessarily give you permission to send it elsewhere. Nor does paying for your own AI subscription make it an approved workplace tool.

For someone else’s confidential information, consider the obligation you have to them. The Solicitors Regulation Authority’s guidance on AI discusses this explicitly for legal professionals. Other professions have their own requirements.

If the task can be done without the private material, use a general question first. Ask how a kind of contract clause works before uploading the contract itself.

2. Send the part that answers the question

You rarely need to upload an entire folder for help with one paragraph.

For a confusing cancellation clause, share the clause and any definitions needed to interpret it. For a spreadsheet formula, build a small example with invented figures that reproduce the problem.

Make clear when you have supplied an excerpt. An assistant should not claim to have assessed the complete document when it has only seen two pages.

Do not cut away information that changes the meaning. A sentence beginning “except where” might contain the most important qualification in the agreement.

The ICO’s minimisation principle is a helpful starting point. Use the information necessary for the purpose, rather than everything available.

3. Inspect the parts you cannot immediately see

Read comments and tracked changes before uploading a working document. A removed paragraph may still be present as a revision.

Spreadsheets can contain hidden sheets or rows. A file may also carry author information or other metadata.

Microsoft’s Document Inspector can help find several kinds of hidden content in supported Office files. Its coverage has limits and differs between applications, so read the guidance for your file type.

Run inspection on the separate copy. Some removal actions can change the document in ways you cannot easily undo.

Exporting to another format is not proof that sensitive content disappeared. Open the actual file you intend to upload and check it.

4. Remove private details properly

Putting a black rectangle over text is not always redaction. The original text may remain underneath and still be selectable.

Use a redaction feature designed to remove the underlying content. Save the result, reopen it, then try searching for the information you intended to remove.

That check is useful, but it is not a guarantee that every hidden element has gone. For particularly sensitive material, use your organisation’s approved process or share a newly written summary instead.

A summary also needs checking. Removing a name while leaving a distinctive incident may still identify the person involved.

5. Limit what the assistant can do with the file

Reading a document and obeying instructions inside it are different things.

An uploaded file can contain malicious instructions aimed at the assistant. The NCSC’s explanation of prompt injection makes clear that this problem cannot be solved completely with a reminder to ignore suspicious text.

For a reading task, avoid granting unrelated action permissions. If the assistant only needs to explain a report, it does not need to send messages on your behalf.

Ask for references to the relevant page or passage, then check them in the original. A confident summary can still miss a condition or misread a table.

6. Keep useful context without uploading the file again

AI Passport holds selected personal context that connected assistants can use with your permission. It can help when a long document contains a small amount of background you repeatedly need.

Suppose a project brief explains your audience and preferred writing style. You could propose a short project memory containing that information, rather than uploading the whole brief every time you change assistants.

Read the proposal carefully. Keep confidential budget figures out if the writing task does not need them. Approving the memory is separate from allowing a connected app to retrieve the relevant category.

Passport gives you controls over that access. It does not make an unsuitable document suitable for another service, and text an assistant receives remains subject to that service’s rules.

Before pressing upload, open the final copy one last time. Check that it contains what you intended to send, rather than relying on what you remember removing.