Can AI Unblur a Photo You Tried to Redact?

‘Unblur’ in research papers usually means deconvolution or learned priors: guess a sharp image that would have produced the blurry measurement. That guess uses surrounding context — edges, textures, language models of what text looks like. It works best when blur is linear, uniform, and shallow relative to pixel size, and when the sensor still captured fine detail that blur merely mixed.

Redaction should assume a motivated viewer with offline tools, not only a casual scroll. The question is not ‘will a random website undo this?’ but ‘did I remove or replace the samples?’

Mild blur: partial recovery is plausible

A Gaussian blur with a small radius on a 4K screenshot of a credit card convolves neighboring pixels. High frequencies are attenuated, not zeroed. Deconvolution algorithms (Richardson–Lucy, Wiener) and modern CNN/restoration models can sharpen edges enough that a determined analyst guesses digits — especially if font and layout constrain the search space. Security research on face anonymization has shown similar limits for soft blur on faces: identity leakage scales with blur strength and resolution.

That is not a claim that any free ‘AI enhance’ button will read your card. It is a claim that soft blur is concealment for casual viewing, not destruction of information.

Pixelation and solid fill: the signal is gone

Pixelation replaces a block of pixels with one average color (or a downsampled cell). Information within the cell is averaged away. Upsampling cannot recover which original pixels contributed without external knowledge. A solid black or white fill is stronger still: you replaced the region with a constant. No blur kernel to invert.

Heavy mosaic (small cell size relative to feature size) pushes toward the destroyed regime. A 4×4 mosaic on a 12-digit number might still leak length and spacing; a 20×20 mosaic on the same region removes readable glyphs. Black box removes even spacing cues inside the box if the box is tight with margin.

What AI upscaling adds to the picture

Super-resolution models hallucinate plausible detail — pores, fabric weave, serif shapes — conditioned on training data. They can make a blurry face look sharper without being the true face. For text redaction that hallucination is dangerous in a different way: invented strokes can look readable even when wrong. For redaction integrity you do not want inference; you want absence.

Commercial ‘enhance’ features on phones use similar ideas for noise and motion blur on camera photos. Screenshot redaction is harsher: binary text on flat background is the easiest case for partial recovery under mild blur.

What HideShot defaults imply

HideShot offers Black Box, Blur, and Pixelate because users want all three aesthetics. For credentials, medical strings, and account numbers, the product warning is honest: prefer Black Box (or heavy Pixelate verified at zoom). Blur is for bystanders in a crowd photo when you accept cosmetic anonymization and inspect the export. That aligns with technical consensus: irreversible redaction removes or overwrites data; reversible obfuscation is not the same control.

How to redact with AI limits in mind

1. Classify the secret: identifier (black box) versus face in background (pixelate/blur after zoom test).
2. Cover with margin. Partial characters leak checksums and guesses.
3. Download PNG. Zoom past 100%. Attempt select/copy if the source was text-like — see verification guide.
4. Do not republish the unmarked original alongside the covered copy.

Mistakes

Light blur on sharp screenshot text because it ‘looks nicer.’ That is the recoverable regime.

Assuming AI cannot help attackers because you tried one demo site. Threat models include offline research tools, not only consumer buttons.

Pixelate so lightly that cell boundaries align with character strokes. Increase cell size or black-box.

Mode depth: blur vs pixelate vs black box. Verification: how to verify a photo was actually redacted.