Redaction vs Blur: Which One Actually Removes the Information

A blur is a filter. For each pixel it computes a weighted average of its neighbors and writes the averaged value. The result is a new image in which the original high-frequency detail has been reduced. Information is degraded, not deleted. Where the original detail was large relative to the filter radius, the degraded result still carries a recognizable shape.

A solid fill is a replacement. The pixels inside the region are assigned a single color value. In the written file there is no averaged remnant of the original — there is a constant. Nothing about the glyph shape, the character count, or the spacing survives, because none of those values were written.

Overlay is a third thing, and it is the one that causes the most embarrassing incidents. Drawing a black rectangle on top of an image without flattening produces a file that contains both the rectangle and the content underneath. Presentation slides, design files, and PDFs all behave this way by default. The redaction looks correct on screen and is empty when anyone looks one layer down.

Blur is most effective when the region contains a lot of detail at a scale much smaller than the filter radius — a dense texture, a face at distance, a page of body text. Averaging destroys that kind of detail reliably.

The failure case is a small number of large characters. A nine-digit Social Security number or a four-digit card tail in a screenshot occupies few pixels per character but each character has strong strokes. Averaging across a stroke produces a gradient at exactly the places where the glyph differs from its neighbors. At sufficient zoom, or with a deconvolution step, the character is often recoverable.

There is a second, cheaper attack that does not need deconvolution. If the attacker knows the field type, they can enumerate candidates and compare. A Social Security number has a restricted format; a card number has a checksum. A candidate set of a few thousand possibilities against a degraded image is a tractable problem in a way that recovering an arbitrary sentence is not.

A blur is defensible when the information underneath is not the thing you are protecting. Blurring a background to make a subject stand out is a composition decision. Blurring a face in a crowd photograph is a reasonable trade-off between privacy and legibility, and it is also a case where a human face has enough redundancy that a low-information blur is genuinely hard to reverse.

A blur is not defensible for credentials, identifiers, names, addresses, or any text a person would be harmed by having read. That includes things that feel minor: an order number that unlocks an account lookup, a workspace name that identifies an employer, a username that links to a public profile.

The working rule is simple. If the region contains a value from which someone can do something — log in, verify, look up, impersonate — use a fill. If the region contains context whose exact form does not matter, a blur is fine.

Pixelation downsamples the region into blocks and paints each block with a single color. It feels more thorough than a blur because the result looks blockier, but the mechanism is the same category: it averages and writes the average. The block grid is a deterministic function of the block size, which means the averaging can be inverted if the block size is known or can be estimated.

Pixelation has one genuine advantage: it clearly communicates that something was intentionally removed, which matters when a document needs to look processed. It has one genuine disadvantage that is worse than a blur in some cases: the block grid lines up with the character pitch in ways that can preserve spacing information.

This tool offers pixelate because some layouts need the surrounding text to remain legible, and it paints an opaque base underneath the mosaic so nothing shows through. But the mosaic itself is built from original pixels at reduced resolution, and the interface says so. For short strings, solid fill is the safer choice.

Open the exported file, not the editor. The editor shows boxes composited over an image in memory; the file is what a recipient receives. Zoom to actual size and look at the edge of each covered region — a one-pixel sliver of a character is the most common real-world failure.

For a PDF, try selecting the text under the black box. If it is selectable, the redaction did not happen. The reliable fix is to rasterize the page to an image and redact the image, because an image has no text layer to leave behind.

For a JPEG, be aware that the format is lossy and that re-encoding can reintroduce artifacts around an edited region. PNG, which this tool exports, is lossless, so what you see in the editor is what gets written.