What Are AI Watermarks?

“AI watermark” is used loosely. People use it for on-image labels, hidden Unicode in copied text, cryptographic manifests attached to files, and statistical patterns that a model’s owner may be able to detect. Those mechanisms are not interchangeable, and they are not equally inspectable by a third party.

Four different ideas that get mixed together

• Visible marks: text, logos, or overlay patterns that a person can see without tools.
• File credentials: signed provenance attached to a file, such as C2PA Content Credentials, which can be present, stripped, or never written.
• Hidden characters: non-printing Unicode copied with text. These can be listed when they exist, but they are not proof of a specific model on their own.
• Statistical or model-side watermarks: patterns designed so a provider (and sometimes a licensed detector) can later estimate whether output came from a particular system. Independent tools may not be able to confirm or remove them.

What an independent inspector can actually do

An independent site can inspect what is in the bytes or characters a user supplies: metadata fields, C2PA manifests when they are present, and Unicode code points that are actually in a string. It cannot honestly claim to “see” a provider’s private detector, and it cannot treat a clean inspection report as proof that the content was not generated by AI.

Removal claims require a verification path

Search queries such as “Claude watermark remover” describe a real user intent: people want to know whether a mark exists and whether it can be taken off. Visible labels and some metadata can be removed in a way a tool can re-check. Invisible statistical watermarks are different. Unless a result can be independently verified, this site will not claim successful removal.

How this site will label claims later

Provider pages and research notes will use a four-level evidence hierarchy: officially documented, independently verified, observed in limited testing, or unverified. Phase 1 publishes the hierarchy and the vocabulary. It does not invent provider watermark statuses.

Related reading on this site

C2PA Content Credentials are a provenance standard, not a synonym for an invisible watermark. Hidden Unicode in copied text is a separate inspection problem. Both are covered in dedicated guides, and both will later connect to the Tools index as those inspectors ship.