Start with the mechanism
What is an AI text watermark?
The phrase can describe several unrelated mechanisms. Treating them as one removable thing creates misleading results: a hidden character is not a statistical signal, a Content Credential is not an authorship score, and none of those is proof that a person did or did not write the text.
Five questions hiding behind one search phrase
Invisible Unicode and copy/paste artifacts are literal scalar values in the text. They may affect spacing, cursor movement, line breaking, display order or matching. Because they are present as characters, this inspector can reveal them and can propose conservative, reviewable deletions for selected low-risk cases.
A statistical generative watermark works differently. It changes how tokens are selected while a model generates text and is detected with a scheme-specific statistical test. Content Credentials are provenance structures. File metadata lives in a container such as PDF or DOCX. An AI-authorship classifier estimates a category from patterns. Those four mechanisms are not removed by deleting zero-width characters.
- Hidden Unicode: inspected here, with selected cleanup only.
- Statistical watermarks such as SynthID Text: not inspected or removed.
- C2PA Content Credentials: structurally checked and protected from transformation.
- DOCX/PDF metadata: unsupported because this launch accepts plain UTF-8 text and Markdown only.
- AI-authorship scores: unsupported and not presented as watermark evidence.
What a finding means
A finding says that a specific Unicode character or copy artifact exists at a specific location. It does not say who inserted it or why. Ordinary software, web pages, word processors, bidirectional writing systems and emoji sequences all use invisible formatting characters for legitimate reasons.
That is why the result uses names such as “zero width space,” “format control” and “copy/paste artifact.” It shows evidence and lets you decide. It never changes a finding into “AI-written,” “human-written” or “detector-safe.”
The useful narrow job
If copied text behaves strangely, a literal character inventory can be useful immediately. You can locate an unexpected scalar, see a visible token in context, review the proposed edit and obtain an exact receipt for every selected deletion.
If your goal is to remove provenance, defeat a detector or guarantee how another system will classify writing, this is not the right tool. The workbench intentionally refuses possible Content Credentials and does not rewrite language to evade another system.
Primary references
These sources describe the external standards or mechanisms discussed above. The property’s cleanup decisions remain narrower than those specifications.