Decide what answer you need before attaching a file
If you need a count of support issues, a whole customer workbook is excessive. If you need help with a failing function, a small made-up example is usually more useful than a repository archive. Work out what information the question requires, who can give permission to share it, and which ChatGPT account your organization allows. If you cannot answer those questions, ask before uploading.
| Decision | When it fits | Next action |
|---|---|---|
| Do not upload | A live secret, authentication export, prohibited data, or material you cannot disclose | Keep the file on your device. Use fake values if you only need help with its format or an error. |
| Ask the person responsible | A confidential draft or sensitive record might be needed, but you do not know whether or where you may share it | Get permission for this task and account before preparing an upload. |
| Make a smaller copy locally | You may use the data, but the file includes extra rows, hidden content, identifiers, or metadata | Export the required excerpt or count, inspect the new file, then decide again. |
| Proceed after a final check | You have permission to use this account, and the reviewed copy contains only what the task needs | Check the filename, content, audience, and destination immediately before sending. |
Do not put a real secret into ChatGPT to ask whether it is safe. Describe the type of item without its value. For a fuller permission check, use the sensitive-data preflight.
Keep live secrets and other people's records out
A password-manager export, private key, seed phrase, recovery-code sheet, session cookie, or bearer token should stay out of a prompt and out of an attachment. Replace it with an unmistakably fake value if you need help understanding a format or error. Do not attach a file you are not allowed to share, even if ChatGPT accepts it.
Medical reports, bank statements, identity documents, student files, employee records, and customer exports need explicit permission and the right account for the job. Removing a name may still leave account numbers, dates, rare events, free-text notes, or combinations that identify a person. An unpublished earnings draft, contract, legal advice, board paper, acquisition plan, or trade secret also needs its owner's permission. If the task can be answered with invented facts, counts, or a short permitted excerpt, use that instead.
Check the file you will actually send
Work on a copy. Cleaning the visible page is only one part of the review: inspect what the format can carry, export a fresh sharing copy, reopen that copy, and check it again. If you cannot inspect a format reliably, send a plain-text description or a newly created minimal file instead of the source container.
Spreadsheets: look past the active tab
A workbook can hold hidden worksheets, rows and columns, comments, author details, named ranges, external links, formulas, and embedded objects. In supported desktop versions of Excel, save a copy and use File > Info > Check for Issues > Inspect Document. Review the results before removing anything: Microsoft warns that deleting hidden sheets or rows can change calculations, and its inspector can find items that it cannot remove for you. Manually review the source tabs, formulas, connections, and free-text cells that matter to the task. Microsoft's Document Inspector guide lists the checks and their limits.
If you only need to count support issues, create a new file containing just the issue category and count. Open that file to confirm that customer rows, notes, formulas, links, and the original workbook name did not travel with it. If the job really needs one row per person, see how to prepare row-level data; replacing names with codes can still leave people identifiable.
Word documents and PDFs: review the editing trail
A clean-looking Word page can still contain tracked changes, comments, hidden text, headers, document properties, or embedded files. Save a copy and use Word's Document Inspector where available; read every result and reinspect the sharing copy. Accepting a revision can also change the document's meaning, so have the owner approve the final text before release. Microsoft's guide notes that the inspector does not find every way content can be hidden, including white text on a white background or an object covered by another object.
A PDF can carry comments, attachments, metadata, hidden layers, or text beneath a visual cover. For a PDF that must be shared, use a suitable redaction and sanitization tool on a copy, then reopen the output and check visible pages, selectable text, search results, attachments, and properties. Adobe documents Acrobat Pro's sanitization controls. If all ChatGPT needs is a clause explanation, a short excerpt you are allowed to share may avoid sending the whole contract.
Photos, screenshots, audio, and video: inspect the surroundings
Zoom into every corner of a screenshot or photo. A browser tab, notification, name badge, background monitor, QR code, boarding-pass barcode, or booking reference can disclose more than the subject in the center. Check the filename and available location or author metadata too. Make a new flattened sharing copy after removing unnecessary areas, then reopen that exact export. A cropped editor view is not the final file check.
Recordings can include other people's voices, faces, names spoken aloud, a shared screen, or confidential discussion outside the segment you wanted analyzed. If the task only needs a summary of one approved passage, prepare a reviewed transcript excerpt locally. Do not upload a whole meeting recording simply to extract a few agreed facts.
Logs, code, and archives: reduce them to a reproducible example
Logs can expose authorization headers, cookies, tokens, customer identifiers, IP addresses, internal URLs, and payloads. Code folders and ZIP files can include .env files, test fixtures, private configuration, keys, and repository history. A search for familiar secret patterns helps, but it cannot find every sensitive value. Read the lines you plan to send, including free-text messages, then create a new excerpt with invented identifiers and only what is needed to explain the failure.
For example, to debug a retry bug, provide a small synthetic function, a fake endpoint such as https://example.invalid, a short sanitized error trace, and the expected result. Keep the repository and full production log local. If a real credential was already disclosed, follow the incident process and revoke or rotate it; changing the text in a later prompt does not undo the earlier exposure.
Run the decision on common files
| What you have | Decision now | What could be shared later |
|---|---|---|
| A password export or log containing a live OAuth token | Do not upload | A fabricated example of the file format or error, with no usable credential |
| A boarding-pass photo with a name, booking reference, and QR code | Do not upload the photo | Type only the flight time or route needed for the question; leave out the booking code and barcode |
| An uninspected sales workbook | Hold and inspect | A count or selected values from a new file, after checking hidden content |
| An unpublished earnings draft from your employer | Ask who may see it and which account to use | Only the exact material your employer permits for this task |
| A named medical report for someone else | Keep it out until the responsible person or organization permits this use | A made-up question; if permission is granted, only the part needed for the task |
| A published manual or deliberately synthetic script | Proceed after checking permission and leftover sensitive details | The relevant excerpt or minimal script, after a final file and destination check |
The screenshot below came from an exercise using descriptions of eight fictional files; no files were uploaded. ChatGPT classified those descriptions, but it did not inspect a hidden spreadsheet tab, image metadata, or a real document. Use the decisions above as a starting point, then check the actual file yourself.
Check where ChatGPT may keep or share the material
On a signed-in personal account, Settings > Data Controls > Improve the model for everyone controls whether new conversations are used to train ChatGPT. Turning it off does not remove a conversation from history, strip identifiers, or grant permission to disclose a file. OpenAI says its business offerings have different default model-improvement treatment; confirm the exact workspace and its administrator's rules before relying on them. OpenAI's business data page describes the default.
A Temporary Chat stays out of history while temporary and is not used for model improvement. It does not create or update memories, though a personalized Temporary Chat can use existing memories. OpenAI may keep a copy for up to 30 days for safety. Saving that chat later makes it a regular chat, with the account's settings and retention rules; eligible uploaded files may then be saved to Library.
Where you attach the file also matters. OpenAI says a file saved in Library can remain after you delete its conversation. Files placed in projects or as knowledge for a custom GPT have their own retention path. A shared project can make contributed context available to its members. If a GPT sends data through an action, the receiving third party follows its own privacy policy, even when the conversation is temporary. Review the destination and recipients before using these features. Project sharing and GPT actions in Temporary Chat are documented separately by OpenAI.
If you already uploaded the wrong thing
- If you sent a usable password, token, recovery code, or private key, follow your incident procedure and revoke or rotate it immediately. Tell the data owner or security team if the material belongs to someone else or your organization.
- Stop sharing the chat, file, or project. If you created a shared conversation link, remove it under Data Controls; that stops future access through the link, but does not erase copies someone already saved. OpenAI's sharing guide explains the limit.
- Review the places where the content was stored: the chat, Library, a project, or a custom GPT. Delete the appropriate copies and record what was sent and when. OpenAI says deletion is scheduled under the applicable retention rules and may have security or legal exceptions; do not describe deletion as an immediate recall. See the current retention policy.
