Three methods dominate: blur, pixelation, and a solid shape over the top. They are not equivalent, and the differences matter more as the stakes rise.
Blur
A Gaussian blur is a convolution - each output pixel is a weighted average of its neighbours. That is a reversible operation in principle, and deconvolution is the standard way to reverse it. In practice recovery depends on how strong the blur was, how much noise the image has and whether the result was re-compressed afterwards.
The rule of thumb is uncomfortable but useful: if you can still see the structure of what is underneath - the rhythm of characters on a plate, the shape of eyes and mouth - then the structure is still in the file, and structure is what a recovery algorithm needs.
Pixelation
Pixelation replaces a block with its average colour, which sounds destructive and often is not. If the blocks are large relative to the content, the information is genuinely gone. If they are small, each block is a measurement, and a grid of measurements over something with a known structure is a solvable problem.
Pixelated text is the well-documented disaster case, because text has a small, known alphabet. Someone can render every candidate string with the same font at the same size, pixelate each one identically, and compare. This is not exotic - it has been demonstrated repeatedly on redacted documents and on plates.
What actually removes the information
A solid fill. A flat bar replaces the pixels. There is no gradient, no average, nothing carrying a trace of the original. It looks blunt, which is exactly why it is trustworthy.
A stamped label. Same property, with the advantage that it announces itself. A reader can see that something was deliberately removed rather than wondering whether the photo was just badly taken.
Cropping. Effective, with one caveat: some cameras and editing pipelines keep an embedded thumbnail generated before the crop. It is rare and it has happened enough to be worth knowing about.
When the weaker options are fine
This is not an argument that everything needs a black bar. A blurred plate on a car listing that will be deleted in three weeks is a reasonable trade against a photo that looks vandalised. The judgement is about permanence and audience.
- Short-lived, small audience, low consequence - blur or pixels at decent strength are fine.
- Permanent, public, or attached to an argument that might attract attention - solid fill.
- Anything involving a document, an ID, a signature or a screen showing account details - solid fill, always.
The step that undoes all of it
Applying a strong cover and then exporting a file that still carries the original in its metadata, or sharing the untouched original by accident alongside the edited one, is the most common real-world failure. It is not a flaw in the method. It is filing.