Claude's invisible text watermark, explained honestly

What Claude does

Claude models released on or after August 2, 2026 mark output in two ways, everywhere Claude is offered:

• Embedded text watermark. An imperceptible statistical watermark woven into the words at the model level. Invisible, doesn't change meaning, survives copy-paste.
• Signed C2PA metadata on files. Generated .svg, .png, and .jpg files carry cryptographically signed C2PA provenance.

It applies across the API, Claude, Claude Code, Claude Cowork, and Claude Tag — and can't be turned off by a paid tier or setting. The driver is EU AI Act Article 50.

What a detected mark does and doesn't prove

A mark means the content may have been processed by Claude. It does not prove authorship. If a human writes 900 words and asks Claude to tighten the prose, the output carries a mark. Conversely, no mark proves nothing: text from older models, heavily rewritten text, or very short passages may carry no detectable signal.

Can I remove the watermark?

Honest answer: not with public tooling today. Anthropic has not shipped a public detector, and a statistical signal in word choice isn't a character you can strip. Tools that claim to "remove Claude's watermark" are selling something that doesn't exist. What is removable is the visible-invisible stuff — the zero-width and hidden Unicode characters that slip into copied text. That's exactly what NoAtMark cleans.

Our stance on "bypass" tools

We don't build watermark-removers or detector-bypass tools. Stripping C2PA provenance or laundering text to hide AI usage fights transparency mechanisms, and the "evade detection" claims are statistically unreliable anyway. We build text hygiene: find and remove invisible characters, and rewrite text to read naturally — not to fool detectors.