How do you remove personal information from a PDF before giving it to an AI?

1. Why this is different from redacting for a person

When you send a redacted PDF to a person, the risk is that they notice the redaction is fake and copy the text out. When you give a PDF to an AI assistant, there is no noticing involved. AI assistants extract the document's text layer — every character the file contains — and many look at an image of each page as well. A rectangle drawn on top of a word changes the image; it does not remove the word from the text the model is given.

People have been caught out by this long before AI. In 2019 journalists recovered the blacked-out passages of a court filing by Paul Manafort's lawyers simply by copying and pasting them (ABA Journal). An AI assistant does the equivalent automatically, every time.

Assume the AI sees everything you could get by pressing Ctrl+A and pasting into a text editor. Because in effect, that is what it does.

2. The three steps

1. Redact, properly. Mark every name, account number, address, reference and signature — including repeats in headers and footers, which are easy to miss. Use a tool that removes what is under the mark. The dependable way turns each page carrying a redaction into an image with the marked areas painted out, so there is no text layer underneath left to read.
2. Strip the metadata. The author field frequently holds a full name, and the organisation, software and dates travel with it. That information is often stored twice, and the second copy — the XMP packet — is the one simple "remove metadata" options miss. What a PDF carries, and where.
3. Check the result. Open the new file, select all, copy, and paste into a plain text editor. Anything you can paste, the AI can read. Then look at the document's properties. This takes thirty seconds and is the only step that proves the first two worked.

3. Placeholders or black boxes?

For an AI, black boxes are often the worse choice. If you want a model to summarise a contract or explain a bank statement, deleting the names leaves it guessing who is who. Replacing them with consistent placeholders keeps the document intelligible while the real values stay on your machine.

4. Also worth checking in the AI tool itself

Redaction controls what you send. The assistant's own settings control what happens to it next. Most AI products let you opt out of your content being used to train future models, and many offer a temporary or incognito chat that is not kept in your history. The names and locations of those settings change often, so look in the product's data or privacy settings rather than relying on a guide — and remember that an opt-out applies to what you send after you set it.

Common questions

How do I remove personal information from a PDF before uploading it to ChatGPT?

Do three things, on your own device, before the upload. First, redact properly: remove names, account numbers and other identifiers with a tool that deletes what is under the mark rather than drawing a box over it. Second, strip the metadata, including the XMP packet, because the author's name is often recorded there. Third, check the result by selecting all the text and pasting it somewhere — anything you can paste, the AI can read. Only then upload.

Is drawing a black box over text in a PDF enough before sharing it with an AI?

No. A rectangle drawn with a markup or shape tool sits on top of the text; the text itself is still in the file. AI assistants extract a PDF's text layer, often alongside an image of the page, so the words underneath the box reach the model even though they are invisible on screen. The content has to be removed, not covered.

Should I replace names with placeholders or black them out?

For an AI, placeholders often work better. Replacing a name with "Client A" or an account number with "ACCOUNT-1" keeps the document readable, so the model can still follow who did what, while the real value never leaves your device. Blacking out is right when the document itself is being sent to a person; placeholders are right when you want an analysis of the content.

Is it safe to use an online redaction tool before uploading to an AI?

Only if it does not upload the file. Many online redaction sites work by sending the document to their server, which means the unredacted original is transmitted to a third party in order to protect it from a different third party. A redaction tool that runs in the browser avoids that: the original is processed on your device, and only the redacted version ever goes anywhere.

Does redacting a PDF make the text unsearchable?

With a redaction method that cannot leak, the redacted pages usually are. The dependable approach turns each page carrying a redaction into an image with the marked areas painted out, which guarantees nothing survives underneath but also means those pages' text can no longer be selected or searched. Pages with no redaction on them can be kept exactly as they were. Run OCR afterwards if the redacted pages need to be searchable again.