You have a 40-page contract and a simple wish: paste it into ChatGPT, Claude or Gemini and ask what it says. The problem is everything else that rides along: the names, the addresses, the account numbers, your signature block, a counterparty’s phone number in the footer. The AI does not need any of it to summarise clause 7. But once pasted, it is in someone else’s system, under someone else’s retention policy.
The fix is obvious: remove the personal data first. The trap is in how.
The trap: "private" redaction that uploads your text
The market has noticed this need, and a wave of tools now offers exactly this, "redact before you paste into an AI." Read their fine print carefully. A common pattern is a tool that promises your file never leaves your device, and then, to find the sensitive text, sends the document’s extracted text to a server-side AI for the detection step.
Think about what that means. The detection pass is the part that reads everything, every name, every number, the whole document, because that is its job. If the detection runs on a server, then the complete sensitive content of your document was uploaded before a single box was drawn. The redaction may be local; the reading was not. For the specific job of keeping a document away from other people’s machines, that is the whole game lost in step one.
So the one question that separates these tools: where does the detection run?
Detection on your device, with a receipt
Our Redact PDF tool answers that question with a measurement instead of a sentence. Both detection layers run in your browser:
- The pattern scan reads the document’s text locally and suggests boxes over the structured data: email addresses, card numbers (checksum-validated), ID and Social Security numbers, IBANs, phone numbers, dates.
- The AI pass for names and postal addresses runs a language model on your device, and it only ever suggests boxes over text that literally appears on the page.
And after each scan, the page shows a live receipt: the actual number of bytes that left your browser during that run, measured by wrapping the network itself, not asserted by a marketing page. It reads 0 B uploaded during this scan, and you can copy that receipt. A tool that sends your text away for detection cannot print that line, because for it the line would be false.
Everything is a suggestion you review. Tick the boxes that matter, dismiss the false alarms, add anything the scanner could not know, and export. On export the covered content is removed from the file, not hidden behind a rectangle.
Then strip what you cannot see
Redaction handles the personal data you can point at. PDFs also leak invisibly: author names in metadata, comments and sticky notes carrying colleagues’ names, form fields still holding typed answers, hidden layers with content you thought was deleted.
That is the second step: run the redacted file through Clean PDF for AI. It strips the metadata and XMP, flattens form fields into the page, removes every annotation and discards hidden layers, leaving a plain copy where what you see is all there is. Redact the visible, then clean the invisible, and the file is genuinely ready to paste.
The two-minute flow
- Open Redact PDF and add your document. It stays in your tab.
- Scan. Review the suggested boxes from the pattern scan, and on a desktop, run the names-and-addresses pass too. Check the receipt: 0 B.
- Export the redacted copy; the covered content is gone from the file.
- Run it through Clean PDF for AI to strip metadata, comments, form data and hidden layers.
- Paste with a clear conscience. The AI gets the document; the people in it stay home.
The short version
Pasting a document into an AI publishes it to someone else’s system, so strip the personal data first. But be picky about the stripping: if the detection step runs on a server, your sensitive text was uploaded before anything was removed, and the privacy was theatre. Insist on detection that runs on your device, and better, insist on seeing it measured. Ours shows you the receipt. Try Redact PDF.
Frequently asked questions
Why should I redact a PDF before pasting it into ChatGPT or Claude?+
Because whatever you paste becomes input to someone else’s system, subject to their retention and their policies. A contract, report or statement usually carries names, addresses, account and ID numbers the AI does not need to do its job. Removing them first means the useful content goes in and the identifying content never leaves your machine.
Do not some tools already offer "AI redaction" for this?+
Yes, and this is exactly where you should read the fine print. Many tools that advertise a private, local experience still send your document’s extracted text to a server to run the PII detection. That means the sensitive text is uploaded before anything is removed, which is the very thing you were trying to avoid. The honest question to ask any such tool: where does the detection run?
Where does PDFMergely’s detection run?+
On your device, and that claim is measured rather than promised. The pattern scan and the AI name-and-address pass both run in your browser, and after each scan a live receipt shows the actual bytes that left the device during the run: 0 B. If anything had been sent, the receipt would say so instead.
Is redacting enough, or do I need to clean the file too?+
Redaction removes the visible personal data. A PDF can also leak invisibly, through metadata, comments, form-field data and hidden layers. For a copy that is safe to paste or share, do both: redact the visible details, then run the file through Clean PDF for AI to strip the hidden ones.