How to Redact a PDF So the Words Are Gone
By PDFwarp · · 7 min read
A black box on a PDF still leaves the words underneath. Here is what actually removes them, and what a rewrite will not keep.
A black rectangle drawn over a name does not delete the name. The words are still in the file, and search, copy, and paste can still pull them out. If you need the words gone, you need a new file that never contained them. AI Redact does that, and it does it in a way most "redact PDF" guides never mention: it throws the original page away and writes a new document where each removed span is the text [REDACTED].
That is useful. It is also lossy. Layout, logos, letterhead, and anything that lived only in an image do not survive. Read this before you run a contract, a medical note, or a court exhibit through it.
Why the black box fails
A PDF page is a set of drawing instructions, not a picture with holes in it. When a markup tool paints a black box, it usually adds a rectangle on top of the text. The text operators underneath are untouched. Open the file, hit Ctrl+F or the Mac equivalent, and the "hidden" name still matches. Select-all and paste into a notes app, and it comes along. Some tools call this redaction. It is a sticker.
Acrobat's real redaction tool is different from its drawing tool: it removes the content and can also strip metadata. A lot of free online tools only draw the sticker, then offer a download that looks redacted and is not. If you are checking someone else's "redacted" PDF, search for a word you know was covered. If the search hits, the box did nothing.
There is a second leak. Comments, form values, and document properties sometimes hold the same name or account number that the page shows. Covering the page and leaving the properties panel alone is the same mistake.
What PDFwarp actually builds
AI Redact is a Pro tool. It is not one of the 19 free tools, and it is not the PDF Editor. You describe what to remove — names, phone numbers, addresses, account numbers, or a pattern in your own words. The tool extracts the text, up to 15,000 characters, and sends that text plus your instructions to xAI. The model is told to replace matches with [REDACTED] and leave the rest of the wording. PDFwarp then builds a new file from that string.
The new file is usually a PDF. If the redacted text uses a script the PDF builder cannot draw, you get a Word file instead. You do not get the original pages with holes burned in them. The source you uploaded is not edited in place. Your own copy on your computer is unchanged, which matters: the rewrite is a sharing copy, not a replacement for the record you still hold.
Because the new file is typeset from the remaining words, copy and search cannot recover a span that was replaced. That is the part a black box does not do. The cost is the page. Columns, headers, signature images, stamps, and the position of a table do not come through. Do not file this output as "the same exhibit, redacted." If a reader needs to see the original page with a bar over one line, this is the wrong tool. Say that in the cover note if you send the file to someone who expects a visual redaction.
Step by step
1. Open AI Redact. It is included with PDFwarp Pro at $9 USD/month. A free account cannot run it.
2. Upload the document. A scan that is only a photo of text has little or no extractable text, so the model never sees the names in the picture. Run AI Scan to Text first if you cannot select the words, then redact the text file you get back — and remember the scan images themselves are a separate problem, covered below.
3. Describe categories, not a list of every string. "Remove all personal names, phone numbers, email addresses, and street addresses" works better than pasting fifty values. The model is matching the extracted text, not reading your mind about a nickname you forgot.
4. Download the new file and read it before you send it. Search the output for a name you expected to disappear. Then search for a word that should have stayed, so you can see whether the model ate a sentence it should have kept.
Limits that change the decision
15,000 characters. Text past that cap is not sent. A long contract can come back incomplete, and the only sign may be a one-line truncation note at the end. If the document is longer than a few pages of dense text, split it with Split PDF, redact the pieces, and read each output. Do not assume page 12 was included because page 1 looked right.
Images are out of scope. A name inside a logo, a scanned stamp, a photo of a driver's license, or a signature that is only an image is not in the extracted text. It is not replaced. It is also not in the new file, because the new file has no images. That is safer than leaving the image in place, and it is not the same as blacking the image out on the original page. If the sensitive part is the picture, this tool will not "see" it and mark it. You still have the original, pictures and all, on your machine.
Misses are normal. Unusual ID formats, nicknames, names split across lines, and international phone numbers get skipped. The product page says to spot-check for that reason. A model rewrite is not a certified redaction pass. If a missed number would be a real harm, a human has to read the output. There is no setting that changes that.
The original layout is gone on purpose. Report-style documents become plain paragraphs with [REDACTED] where a span was removed. If you needed the table grid, rebuild the few cells that matter, or keep the original offline and send only the rewritten text.
What to do with the original
The upload is processed on PDFwarp's servers and deleted as soon as your download completes. PDFwarp does not keep a cloud copy of the file, and it does not keep the document text after the request. The text that was sent to xAI is a separate thing: xAI does not train on API inputs without permission, and by default it may keep API requests for up to 30 days for abuse checks. Full detail is in the Privacy Policy.
That deletion does not delete the file on your laptop. If the original still has the names, anyone with that file still has the names. Store it somewhere that matches how sensitive it is, and do not email it "just in case" alongside the redacted copy.
If the document is legally privileged or your workplace requires on-device handling only, do not upload it. A short retention window is not the same as the file never leaving the machine. Are online PDF tools safe? walks through that choice.
What this is not
It is not password protection. A password stops a stranger opening the file. It does not remove a name from the text. Use Protect PDF only when you still want the full document and you have a separate way to send the password.
It is not unlocking. Unlocking removes a password you already know. It does not hide content.
It is not a promise that every identifier is gone. The output is as good as the extracted text plus the spot-check you actually do.
FAQ
Can I undo a redaction in the new file?
No. The removed words are not stored under a box in that file. You cannot peel them off. Your original file, if you still have it, is the undo.
Will the PDF look like the one I uploaded?
No. You get a new typeset document. Headings may remain as text. Page geometry will not.
Does a free account include this?
No. AI Redact is Pro only, with the other AI tools, at $9 USD per month.
The file is a photo of a form. Will this work?
Not on the photo itself. Text that cannot be selected is not extracted, so it is not sent and not replaced. Use a text layer first, then read the result, and do not treat the photo as redacted.