Is pixelation safer than blur?

Not in any way you can rely on. Both keep an average of the original pixels, and both have been reversed in published research and open-source tools. Depix and Unredacter specifically target pixelation.

Does a stronger blur make it safe?

It makes recovery harder, not impossible, and you can't measure how much information is left by looking at it. For anything that matters, use a solid box.

Can AI "unblur" an image?

AI upscalers invent plausible detail, so their output can look sharp and still be wrong. The bigger risk is the guess-and-compare approach used by Depix and Unredacter, which checks candidate text against the blurred image instead of inventing it.

Is it safe to blur a password or API key in a screenshot?

No. Short secrets are exactly the kind of text these attacks recover best. Use a solid box, and if a real secret was ever posted blurred, rotate it.

Is a black box always safe?

Only if it is burned into the image you share. A box saved as a separate layer or as a PDF annotation can be removed, and some screenshot tools have left the original pixels in the file. Export a flattened copy and check it.

Is blurring faces fine?

For a background crowd where nobody is at risk, usually. Research has shown that neural networks can identify blurred and pixelated faces from a known group of people, so if someone's safety depends on not being recognised, cover the face with a solid shape.

Can blurred or pixelated text be recovered?

By RedactSnap · 19 October 2026

Often, yes. Blur and pixelation scramble text, but they don't delete it. Researchers and open-source tools have recovered words, numbers and passwords from blurred and pixelated screenshots, especially short text in a common font. If the text matters, cover it with a solid box instead.

Why blur and pixelation leak information

Pixelation splits an area into square blocks and paints each block with the average colour of the pixels inside it. Blur replaces every pixel with a weighted average of its neighbours. Neither one adds randomness. The same text, in the same font, size and position, always produces the same blocks or the same smear.

That makes them reversible by guessing. An attacker who can see the surrounding unredacted text knows the font and size. They render candidate strings, pixelate or blur them the same way, and keep the candidate whose result matches. Short text with a small set of possible characters, such as card digits, phone numbers, dates, email addresses and short passwords, is the easiest to guess.

Black boxes work differently. A solid box replaces the pixels with one flat colour, so nothing about the text underneath is left in the image to compare against.

What has been recovered so far

This isn't a theoretical risk. Recovery has been shown many times over the past two decades:

2007: German police (the BKA) reversed a "swirl" distortion that a man had applied to his face in photos posted online. Interpol published the restored image, and he was identified and arrested in Thailand. A swirl is not a blur, but it is the same mistake: a distortion that keeps the original information.
2016: researchers at UC San Diego (Hill, Zhou, Saul and Shacham) used hidden Markov models, the technique behind early speech recognition, to read mosaiced and blurred text, including long strings and real redacted screenshots found online. They concluded that neither method is a viable way to redact text.
2016: researchers at UT Austin and Cornell Tech (McPherson, Shokri and Shmatikov) trained ordinary image-recognition neural networks to identify faces and read handwritten digits through pixelation and YouTube's face blur.
2020: Depix, an open-source tool by security consultant Sipke Mellema, recovered passwords from pixelated screenshots by matching blocks against a rendered sequence of known characters. He built it after seeing a company pixelate a password in an internal document.
2022: Dan Petro at Bishop Fox released Unredacter after solving a public challenge to read a pixelated phrase. His advice: use black bars over the whole text, with no pixelation, blurring, fuzzing or swirling.

When recovery is harder, and why that isn't enough

Each published attack has limits. Depix only works on one kind of pixelation (a simple box filter) and needs a matching font. Heavy blur, large blocks relative to the text, JPEG compression and resizing all add noise that makes matching harder. Long random strings in an unusual font are harder to guess than a phone number.

None of that makes blur safe. You can't tell from looking at your own blurred image how much information it still holds, the attacker only needs one good guess, and the tools keep improving. A screenshot you post today can be tried again with next year's tools. A solid box doesn't depend on any of this.

Other ways redactions in images fail

A solid box only works if the file you share really is just the image you see. Watch for these:

Edit layers and undo history. Image editors and some screenshot tools save boxes as a separate layer or keep the original pixels in the file. Export or flatten to a plain PNG or JPG before sharing.
Cropping bugs. In 2023, the "aCropalypse" bug (CVE-2023-21036) showed that Google Pixel's Markup tool had been leaving the uncropped original data at the end of edited screenshot files since 2018. Windows 11's Snipping Tool had a similar problem. One of the researchers recovered his home address from a cropped screenshot of an order email.
Partial coverage. A box that misses the top of a tall letter or the tail of a "g" can be enough to guess a short word. Leave a margin around the text.
Length and context. A box that is exactly as wide as the word reveals its length, and the text around it may give it away. Where it matters, cover the whole line or the whole field.
Metadata. Photos carry EXIF data such as GPS location and the device model, which blur and boxes don't touch.
PDF annotations. A black rectangle drawn in a PDF editor is often an annotation on top of the text, which is still in the file and can be copied out. Use a tool that removes the text or flattens the page.

How to hide text in a screenshot properly

Cover the text with an opaque, solid box, with a margin around it.
Export a new, flattened image instead of saving over the original with layers.
Open the exported file and zoom in on each box to check the edges.
If the image is a photo, make sure the exported file doesn't carry the original's EXIF data.
If a password, API key or token was ever visible in a shared image, rotate it. Redacting it afterwards doesn't undo the leak.

How RedactSnap handles this

RedactSnap's redaction tool opens PNG and JPG images as well as PDFs and draws solid boxes only. The download is a new PNG redrawn from pixels with the boxes burned in, so there are no layers, no hidden original and no EXIF data from the source file. It runs in your browser and the image isn't uploaded.

Screenshot Studio finds API keys, emails, phone numbers, card numbers and names in screenshots automatically. It offers blur, pixelate and blackout, and blur is the default. For anything sensitive, switch Studio to Blackout before you download.

Redact an image with solid boxes →Open Screenshot Studio