A scanned page is a picture. A tool that searches for text finds nothing to redact and reports a clean result. To redact a scan, a tool must recognise the words and their positions, cover those pixels, and rebuild the text layer.
Why a scan defeats an ordinary tool
Open a scanned contract and try to select a word. You cannot. There is no text, only pixels shaped like letters.
- Text search finds nothing. With no text to search, the tool reports that it found nothing.
- Black boxes drawn by hand may be an overlay rather than a deletion, and you have to check every page by eye.
- The file can still leak through metadata, or through a text layer someone attached earlier.
This is the most dangerous failure in redaction. Nothing looks wrong, so you trust the result and nobody checks.
Adobe and Presidio treat images separately
Adobe's documentation tells you to run Recognize Text on a scanned PDF before you can search or edit it. Acrobat asks you to convert the image first.
Presidio, the best-known open-source tool for finding personal data, ships image redaction as a separate component from its text analyser.
Both treat an image as a separate job. The question is whether you must remember the step, or the tool does it. A step you must remember gets skipped.
How a tool should redact a scan
- Render the page at full resolution. Our pipeline renders at 300 dpi, because a coarser screen preview puts the boxes in the wrong places.
- Recognise the words and their boxes. Without a box for each word, you know a name is on the page but not where to cover.
- Find the personal data with the same layered detection as any document: fixed patterns, a model that spots names, then a check that removes false hits.
- Cover the pixels and rebuild the text layer. The image regions themselves are painted over. An invisible text layer goes back on top, so the document stays searchable and codes can be decoded.
Skip the last step and the scan becomes a flat picture nobody can search. Do the middle steps carelessly and you miss data or black out half the page.
Text PDFs follow the same rule: a mark is not a redaction unless the content beneath is gone. See redacting legal PDFs without the cloud.
Three ways a scan fails quietly
- Wrong language data. Eight language packs ship: English, French, German, Spanish, Italian, Dutch, Portuguese and Polish. Each is paired with English, which is better at reading email addresses and account numbers. A scan in another language still produces words that look real and are wrong.
- A shape instead of a deletion. A black rectangle drawn on an image can be moved in another viewer. Covering the pixels changes the picture itself.
- An old text layer. If a copier or an earlier tool added a text layer, covering the image leaves it in place. The text comes back with a copy and paste.
Recognition is never complete. The tool flags what it finds, you make the call, and the file stays locked until you confirm. No local detector is exhaustive on a scan, and one missed name is enough.
Why on-device matters most here
Scans are often signed agreements, medical records, ID copies and handwritten notes. Uploading them to a cloud service for redaction is the highest-stakes case. Uploading them is what exposes them.
On your own machine, nobody else has to secure the scan. There is also a browser version. The document stays in the tab, though the app sends an anonymous notice when it starts. The document never leaves your device in either case.
Frequently asked questions
Why did my redaction tool say no personal data was found in a scanned PDF?
It searched for text, and the scan has none. You need a tool that recognises the text in the image itself.
Will redacting a scan make it unsearchable?
Only if the tool does not rebuild a text layer. A complete pipeline writes one back, without the removed data. That layer also lets you decode a reversible code later.
Can the removed text still be recovered from a redacted scan?
Not if the pixels were covered and no hidden layer kept the text. Painting over the image while an older text layer remains leaves a file that looks redacted. A copy and paste still returns the original.
Does this work for photos and phone pictures of documents?
Yes, a photo is an image like any scan. Phone pictures are more often skewed, shadowed or cropped, and recognition suffers. Flatten the page, use even light, and review what the tool found.
In short
If a tool cannot read text in an image, it cannot redact a scan. promptShield runs recognition, detection and covering on your own device, for scanned PDFs and images. Try it on your own document, or read why we built this part ourselves.