Watermarks: Visible and Invisible | AI Learning Guide

Watermarks split into two kinds, and the difference matters for how much weight you give them. Visible watermarks are the overlays you can see: a corner logo, a repeated pattern, a text credit. They are easy to spot and just as easy to crop out, so seeing one tells you something, but not seeing one tells you almost nothing. Invisible watermarks are embedded in the pixel data and are not meant to be seen. SynthID is a well-known example: it alters pixel values in a way designed to survive cropping, compression, and resizing, and you need the matching verification tool to detect it. The limits are the same as for metadata. Heavy re-encoding, aggressive cropping, or deliberate removal tools can degrade an invisible mark. And a surviving watermark proves only that a particular tool touched the file, not that the content is true. A missing watermark proves nothing, because most media never had one.

0:00 / 0:00

Auto-play

Two kinds of watermark

Visible

Overlay logo, repeated pattern, or text credit
Spotted by eye, no tool needed
Removed by cropping or painting over
Presence is informative; absence means little

Invisible

Embedded in pixel or signal data, e.g. SynthID
Designed to survive cropping, compression, resizing
Requires the matching verification tool to detect
Degraded by heavy re-encoding or removal tools

What a watermark does and does not prove

A surviving watermark proves that a particular tool or pipeline touched the file. It does not prove the content is true or accurate. A missing watermark proves nothing at all, because most media never carried one.

References

[1]SynthID: Identifying AI-generated content with SynthIDdeepmind.google