Watermark robustness measures whether the signal survives the kinds of changes expected after content is generated. A statistical text watermark may tolerate small edits because its pattern is distributed across many token choices, while extensive rewriting can remove enough of the pattern to weaken detection.
Robustness must be tested together with false-positive behavior and content quality. A stronger signal may be easier to detect after modification but could distort output more, and no robustness result proves that every possible transformation or unmarked model will be detected.
ELI5
Watermark robustness describes how well a hidden mark survives when content is changed. A robust watermark can still be detected after some ordinary edits, while a fragile one disappears easily.
For example, correcting a few words may leave a statistical text watermark intact because the pattern is spread across the passage. Rewriting most sentences can weaken it, so the absence of a mark does not prove that no AI was involved.
