Is Blurring Text in a Screenshot Actually Safe?

A Gaussian blur replaces each pixel with a weighted average of its neighborhood. The weights are highest at the center and fall off with distance. Two things follow from that: the result preserves local brightness, and it preserves the boundary between light and dark regions at a scale larger than the radius.

For a character with strong vertical strokes — a 1, a 7, an I, an l — the light-dark boundary is exactly what the blur preserves. Increasing the radius widens the transition but does not eliminate the position of the edge. At sufficient zoom the strokes remain locatable, and the relative positions of strokes are most of what distinguishes one character from another.

Recovering an arbitrary sentence from a blurred image is hard, because the space of possible sentences is unbounded. Recovering a structured value is a different problem, because the space of possible values is small and constrained.

A Social Security number has a fixed shape and a restricted range in each group. A card number has a checksum: for any fifteen digits there is exactly one valid sixteenth. A phone number has a known number of digits and a valid area-code list. In each case, a partially readable blur narrows the candidate set from millions to something a script can enumerate in seconds.

That is the argument against blurring credentials specifically. The blur does not need to be fully reversed; it only needs to leak enough to prune. A degraded image of a random phrase is a puzzle with no answer key. A degraded image of a checksummed number is a puzzle with a verifier built in.

Faces in a crowd, backgrounds behind a subject, license plates on a distant vehicle, and screen content behind a speaker are all reasonable blur targets. In each case the information is not a value someone can act on directly, and the visual redundancy is high enough that a partial reconstruction does not resolve to a specific identity or credential.

Even here, the standard is not 'blurred therefore safe'. It is 'the residual information does not enable an action I would object to'. A face blurred with a two-pixel radius in a high-resolution photo is not meaningfully anonymized; the same face at low resolution with a large radius is.

Use a solid fill for anything a person could act on: names, addresses, account numbers, card digits, identifiers, email addresses, phone numbers, tokens, and keys. A fill is not a stronger blur; it is a different operation that removes the pixels instead of averaging them.

If the concern is that a solid block looks visually intrusive, the alternative is a fill in a color that matches the surrounding background. It reads as a design element and still removes the content. What is not an alternative is reducing the blur radius until the block looks less obvious — the block's visibility is the point.

Finally, verify in the exported file rather than the editor. Compositing in a live editor can differ from what a flattened export produces, particularly around antialiased edges.