Claude Watermark: The Straight Version

The mechanism, in one paragraph

At every step a language model has several words that would work equally well. Anthropic's watermark uses a secret key, plus the words immediately before, to decide which of those equally good options gets picked. A reader cannot see the difference. A verifier holding the key can measure that the pattern of picks is not what chance would produce.

Three things this is not

It is not a hidden character. Anthropic's second design principle is exact: "Nothing is added to the text and there are no hidden characters."

It is not an identifier. Anthropic states that the mark "carries no identifying information and can't be traced to a specific person, organization, or chat."

It is not a claim about authorship. In Anthropic's framing, "A watermark only helps test whether Claude might have produced or processed the content. It doesn't say anything about ownership or authorship." Proofreading a human draft through Claude can plant the mark on writing a person wrote first.

Where it applies

The rule is the model's launch date. Models launched on or after August 2, 2026 support machine-readable marking at launch, and older models are being retrofitted against an EU deadline of December 2, 2026. Marking covers Claude, the API, Claude Code, Cowork and Claude Tag, including Claude reached through AWS, Google Cloud and Microsoft Foundry.

Inside code, the mark applies where wording is a genuine choice, such as comments and error strings, and is skipped where an exact token is required for the code to work.

Files are a separate mechanism

Text carries the statistical mark. Files that Claude generates carry signed C2PA provenance metadata instead, which is an ordinary part of the file. That metadata can be read locally, and it can be removed locally. The text mark cannot.