Why pixelation fails
Pixelation — the mosaic filter — replaces each block of pixels with their average colour. That average is not random: it is a deterministic function of what was underneath, which means it carries information about the original content.
If an attacker knows the font, size and rendering of the original text, they can reconstruct it by brute force. Render every candidate string, apply the identical mosaic filter, and compare the resulting blocks against the redacted image. The match that fits is the original text. This is not theoretical — it is a published, implemented technique, and tools that do it automatically have existed for years.
The attack is strongest exactly where redaction matters most. Content with a small, known character set and predictable formatting — account numbers, card numbers, dates, IP addresses, sums of money — has a small enough search space to exhaust quickly. Redacted digits in a screenshot of a familiar interface are close to the ideal case for the attacker.
Why blurring fails too
Gaussian blur is a convolution: each output pixel is a weighted average of its neighbours. Like pixelation, it is deterministic and information-preserving in the mathematical sense — the operation has an inverse, and deconvolution can recover a usable approximation of the original when the blur radius is modest.
Even where exact recovery is not possible, the same brute-force approach works: blur a candidate rendering with the same radius and compare. Blurred text with a known font is not much safer than pixelated text.
Heavy blur with a large radius genuinely does destroy information, and at some radius the content is unrecoverable. The problem is that you cannot tell from looking at it whether you crossed that line. 'It looks blurry enough to me' is not a security property.
The swirl and other reversible transforms
Worth a specific mention because it produced one of the best-known cases: geometric distortion filters, like the swirl effect, are fully reversible. They rearrange pixels rather than destroying them, so applying the inverse transform reconstructs the original exactly.
A person who obscured his face with a swirl filter in photographs was identified after investigators simply un-swirled the images. Nothing had been removed — only moved.
The general principle: if the operation is a rearrangement or a smooth averaging, assume it is reversible. Only operations that genuinely discard information are safe.
What actually works
There are three reliable methods, in increasing order of certainty.
Draw solid opaque shapes over the content. A filled black or white rectangle replaces the pixels entirely — there is no underlying data left to recover, because the original values are simply gone. This is the standard method and it works, with one critical condition covered in the next section.
Crop the sensitive region out of the image entirely. If the pixels are not in the file, nothing can recover them. Where the layout allows it, this is stronger than covering, because it removes any question about layering.
Retype rather than redact. For the highest-stakes cases, do not screenshot the real thing at all. Replace the sensitive values with plausible fake ones in the source — change the account number in the page, then screenshot — or rebuild the example with dummy data. This is the only approach with no failure mode, and it is what you should do for anything genuinely serious. It is worth saying plainly that this includes not trusting our own redaction tool: a tool can only cover pixels, and covering pixels is the second-best answer.
The layering trap
Solid shapes only work if they are actually merged into the image. This is where most modern redaction failures now happen, and they happen in documents more often than images.
Drawing a black rectangle in a PDF editor, a word processor or a design tool typically adds an object on top of a layer that still contains the original text. The text remains selectable, searchable and extractable — it is merely hidden behind an opaque shape. Copy the region and paste it elsewhere and the redacted content comes straight out. Government agencies, law firms and newspapers have all published documents with this exact flaw.
The defence is flattening. Export to a flat raster image, or use a redaction function that explicitly removes the underlying content rather than covering it. A useful verification: open the finished file, select all, copy, and paste into a text editor. If the redacted text appears, you have not redacted anything.
What people forget to redact
Even a correctly executed redaction fails if it misses something. The commonly overlooked items, in rough order of frequency:
- Browser tabs and window titles, which often contain names, account identifiers or document titles.
- The URL bar — query strings routinely carry session tokens, email addresses and record IDs.
- Notification banners that appeared mid-screenshot.
- Autocomplete dropdowns showing previous entries.
- The taskbar or dock, showing which applications are open.
- Reflections in glossy surfaces, screens and eyeglasses within the photo.
- File metadata — the screenshot's own EXIF, and the filename itself.
- The embedded thumbnail, which in some editors preserves the pre-edit image inside the file.