Redact a screenshot

A screenshot you have just taken is already on your clipboard. Pressing ⌘V or Ctrl+V on this page opens it here directly, which means the uncovered version never becomes a file on your device that you then have to remember to delete. That is a small thing that turns out to matter: most accidental disclosures of a screenshot are not the redacted copy going astray, they are the original one still sitting in a downloads folder.

Crisp synthetic type is the single worst case for a blur and the single best case for a bar. Everything about a screenshot that makes it easy to read makes a softened version of it easy to read too.

• The obvious field — the balance, the account number, the address — which is why you came.
• The other places the same value appears: a page title, a browser tab, a notification that slid in at the top, the sum at the bottom of a list.
• Names in a sidebar, a thread list or an avatar tooltip. A conversation is identifiable from the participants long before it is identifiable from the message.
• The URL bar, which frequently carries an account id, a session reference or a customer number.
• The clock and the battery, if the point of the screenshot is that you cannot be placed at a particular time.

The tool counts your regions above the canvas and says the number out loud for a screen reader, because a canvas is opaque to anything that is not looking at it, and a silent count is not an inventory.

Everything else in a screenshot has a hard edge — buttons, rules, table borders, the type itself. A softened covering is the only thing on the page that does not, which makes it the thing an eye is drawn to and the thing a reader spends time studying. A flat bar with a clean edge reads as a deliberate part of the document, which is exactly what it is.

There is a practical reason too. A feathered edge is a gradient between the covering and the original, and a gradient over a character is a partial view of that character. It is a small leak, but it is a leak that exists for aesthetic reasons alone.

If you want the full argument about which covering to use where, it is in the guide on reading blurred text.

Why is a screenshot different from a photograph?

A photograph is noisy. Grain, focus falloff, uneven light and lens softness all add randomness that a recovery attempt has to model. A screenshot has none of that: the characters were drawn by a font renderer at a known size on a flat ground, every instance of a letter is pixel-identical, and the picture is effectively noise-free. That is a much smaller problem to solve.

Is there anything special about numbers?

Yes, and it is the reason to prefer a bar over anything softer. A field with a known format — sixteen digits, a sort code, a date, a reference with a checksum — narrows the answer set enormously before a single pixel is examined. Low-entropy text is the easiest thing in this subject to recover and the easiest thing to fill.

I cropped the sensitive part out instead.

A crop is a good answer when it works, because the pixels genuinely are not in the file. The trap is cropping in an app that writes the original back into the file as a thumbnail or an undo record. Saving a covered copy here re-draws the picture from scratch, so there is no second version inside it.

Does the covering survive being sent through a chat app?

Yes. The covering is not metadata and not a layer — it is the pixels. Anything that re-compresses the picture is re-compressing a picture that already has bars in it.