Guide
Is it safe to paste a bank statement into ChatGPT?
Not as it comes. The statement is useful to the AI. The name, address, and account numbers printed on it are not.
Take those out first and you keep the help without the exposure. The rest of this page is how to do that, and how to decide when it matters.
What actually happens to what you paste
Two different things can go wrong, and people usually worry about the smaller one.
1. Your conversation may train the model
On personal ChatGPT plans, which means Free, Plus, and Pro, OpenAI's help pages say data sharing is on by default and your conversations can be used to improve its models. You can switch it off in Settings → Data Controls → Improve the model for everyone. Business tiers behave differently: OpenAI says it does not train on inputs or outputs from ChatGPT Team, ChatGPT Enterprise, or the API by default. A Temporary Chat is also excluded from training and from your history.
Other assistants have their own defaults and their own settings pages. Check the equivalent page for Claude, Gemini, or Copilot rather than assuming they match.
2. The document leaves your computer, permanently
This is the bigger one, and no setting turns it off. Even with training disabled, the file has been transmitted to and stored on someone else's servers. It's now subject to their retention periods, their staff access rules, their subpoenas, and their breaches. You cannot un-send it.
Training is a policy question that can change. Custody is a fact that doesn't.
The part people underestimate
A bank statement isn't one secret. It's a bundle of them, and the bundle is worth far more than any single line:
- Your full legal name and home address, as your bank has them on file.
- The account number, plus the BSB, sort code, routing number, or IBAN.
- A customer reference that identifies you to that bank.
- Your employer, visible from the salary credit.
- Your landlord or mortgage lender, your insurer, your gym, your pharmacy.
Any one of those is ordinary. Together they're enough for someone to pass an identity check while pretending to be you. That's what makes a statement different from a spreadsheet of numbers.
There's a quieter version of the same risk at work. If you're on a managed laptop, or signed into a browser that syncs, or you paste into a shared workspace, the audience for that statement is larger than the chat window suggests.
What the AI actually needs
Think about the question you're asking. For almost all of them the identity is dead weight:
- "Categorise these expenses" needs dates, amounts, and merchant names.
- "Why did my balance drop in March?" needs the transaction rows.
- "Is this fee normal?" needs the fee line and the product name.
- "Help me with my tax return" needs the numbers and their categories.
None of that improves when the model also knows who you are and where you live. Removing your identity almost never costs you answer quality, which is why it's such a cheap precaution.
Three ways to do it
Retype only what you need
Free, and the safest option. Copy the handful of rows into a plain message and leave the header off. Fine for ten transactions, miserable for a year of them.
Redact a copy of the file, properly
Works when you want to hand over the whole document. The word "properly" is doing real work there. Drawing a black rectangle over text in a PDF viewer hides it from your eyes and leaves it in the file, where anyone can select it or search for it. Our companion guide walks through doing it correctly on a Mac and, more importantly, how to verify it worked.
Use a tool that does it before the file leaves your machine
This is the category our own app sits in, so treat the next sentence with the scepticism it deserves. Cloaksmith finds the names, addresses, and account numbers in a document and removes them on your Mac, with no upload, then hands you a cleaned copy you can share. There are other tools that do parts of this. The thing to check in any of them is whether the redaction removes the underlying text or only covers it, and whether the file goes to a server to be processed.
Things that feel safe but aren't
- A black box drawn with markup tools. The text is still there, underneath. This is how redaction failures make the news.
- Telling the model to ignore the personal details. The data has already been transmitted. Instructions don't recall it.
- Renaming the file. The contents and the document metadata are untouched.
- Exporting to a new PDF after annotating. Sometimes it flattens, sometimes it doesn't. Don't trust it without checking.
- Assuming a screenshot is enough. A screenshot of a covered area genuinely does drop the text layer, but it captures whatever else was on screen, including the account number in the page header you forgot about.
A 30-second checklist
- Does the AI need to know who this belongs to? Usually no.
- Remove name, address, account and routing numbers, and customer reference.
- Check the header, the footer, and every page, not just the first.
- Search the finished file for your surname and the last four digits of the account. Nothing should match.
- Then paste.
Cloaksmith does this step for you
It removes personal details from PDFs, Word files, and scans entirely on your Mac, before anything is shared. No account, no cloud, no subscription. First 5 documents free, then $24.99 once. No tool catches everything, so you always review the result before you share it.