Redact PDF

About Redact PDF

Saying you redacted something in a PDF can mean two things. Either you drew a black rectangle on top of it, or you removed the underlying data from the file. The first only hides it on screen: select the area, copy, paste, and the original text comes right back. This tool does the second.

You mark areas by dragging on the preview or by searching with ready-made patterns (US Social Security numbers, phone numbers, email addresses, 16-digit card numbers) or your own regular expression. There are two ways to apply them. Remove content parses the drawing instructions on each page, deletes only the text, images and vector shapes under your boxes, and keeps the rest of the text searchable. Convert to image renders each affected page with the boxes filled in and replaces the page content with that single picture.

How black boxes fail: the Manafort filing

In January 2019, lawyers for Paul Manafort filed a response in federal court in which the blacked-out passages could be read by copying and pasting them. That is how it first became public that Manafort was accused of sharing 2016 Trump campaign polling data with his former business associate Konstantin Kilimnik, and that the two had met in Madrid. Reporting at the time attributed it to black shapes drawn over text that was never removed.

It is a recurring problem; the American Bar Association's Judges' Journal has collected a series of such failures. The pattern repeats: black highlighter in Word, rectangle annotations in a PDF editor, or redaction marks that were placed but never applied.

What Remove content actually deletes

The tool walks the page's content stream instruction by instruction, tracking the coordinate transforms and font widths, so it knows where each drawing operation lands on the page. Anything that intersects a box is handled like this:

• Text: when glyph widths are known exactly, only the glyphs that overlap a box are replaced with blank spacing of the same width and the rest of the string stays; when they are not, the whole overlapping string is replaced. Either way, text later on the same line stays where it was.
• Images: pixels under the box are repainted in the mask colour. If the image format cannot be decoded, the whole image is blanked. Soft masks (transparency) are made opaque in the same region.
• Vector paths and inline images: shapes crossing the box are cut. A plain rectangle that fully contains the box, such as a background or table cell, is kept.
• Annotations (comments, links, form widgets): removed if they intersect a box, and the form field value is taken out of the field list.
• Hidden text: ActualText and alternate text in tagged PDFs, which never show on screen but do come out in extraction, are checked too.
• The whole file: optionally removes the Info dictionary and XMP metadata, writes a fresh document ID, drops every object nothing refers to any more (including leftovers from earlier incremental saves) and rewrites the file from scratch.

Why more can disappear than you boxed

PDF text is drawn as strings, not individual letters. Some programs write one word per string, others a whole line. With a standard-14 font or a font that carries a width table (/Widths, or /W for Identity-H), the tool splits the string into glyphs and removes only those under the box, so a line stored as one string keeps the name and address and loses just the Social Security number. Glyphs straddling the box edge and glyphs within 10% of the font size of it (usually the spaces around it) go too. If even one width has to be guessed, the whole overlapping string is removed as before. That errs on the safe side, but check the result preview.

When font information is incomplete, for example some CID fonts where the code length and glyph widths cannot be determined, the tool assumes the widest possible glyph for every byte. That too widens the removal rather than narrowing it.

Places the tool does not clean

Bookmark (outline) titles and form field values located away from your boxes can contain the same information. “Also remove the covered text from bookmark titles and form values” (on by default) reads the text under your boxes and replaces the same text (ignoring spacing differences) with ■ in bookmark titles and form values, and has viewers redraw the changed fields. It cannot find text shorter than two characters or written differently (a number without hyphens, say), so when it is off or finds nothing you get the old warning instead. File attachments, text positioned outside the visible page, and the same details repeated on another page all remain unless you box them.

Pattern search works only on the text layer. In scanned documents the letters are pixels, so you have to draw the boxes yourself. Numbers split across two lines, unusual spacing, or digits spread across table cells can slip past. After running, the tool searches the output for the same patterns again and reports how many matches remain, also checking with whitespace removed.

The box itself carries information. The width of a blacked-out name hints at its length, and against a short list of possible people that may be enough to identify someone. For sensitive material, draw boxes that are wider than the text.

When to use Convert to image

If a PDF has unusual font structures and the content removal result looks wrong, or the document is sensitive enough that nothing may survive, choose Convert to image. Pages are rendered with the boxes filled and replaced by a single image each, so there is no text layer left to leak. The cost: no text search, no copying, no screen reader support, larger files and softer text when zoomed in. Government and public bodies often have accessibility obligations for published documents, so check before using it on anything you publish.

Worked examples

How to use it

1. Choose a PDF, then drag boxes on the preview or search with patterns.
2. Review the boxes page by page and choose Remove content or Convert to image.
3. Run it, check the preview and the remaining-matches count, and read any bookmark or form notes.
4. Keep the original somewhere safe and share only the output.

Key terms

Redaction

Deleting sensitive information so it cannot be recovered, as opposed to covering it up.

Content stream

The list of instructions that paints text, images and shapes on a page. This is what the tool parses and rewrites.

Incremental save

Saving changes by appending them to the end of the file, which leaves earlier content inside it.

Frequently asked questions