In brief
Claude text watermarking is a machine-readable marking system for output from supported models. Anthropic says the mark is imperceptible, travels with copied text, and may survive some editing.
Anthropic has not publicly explained the encoding method or released its independent detection procedure. Claims about hidden characters, token biasing, or cryptographic text signatures therefore go beyond the published evidence.
What Anthropic announced
Anthropic signed the EU AI Act Article 50(2) Code of Practice on Transparency of AI-Generated Content and published a plan for machine-readable marking in Claude output.
The plan covers two complementary forms of marking: embedded watermarks in generated text and signed provenance metadata for supported file types. Anthropic says more detailed technical guidance will follow.
What an embedded text watermark means
Embedded means the mark belongs to the generated text rather than a surrounding interface, account record, or downloadable file container. The mark can therefore remain relevant after the words leave Claude’s interface.
Machine-readable means it is intended for a compatible detection process, not ordinary visual inspection. Anthropic has not disclosed the representation, so the phrase should not be treated as evidence for any particular hidden mechanism.
Is Claude's watermark visible?
No. Anthropic calls the text watermark imperceptible and says it does not change the meaning, quality, or readability of Claude’s response.
That distinguishes it from a visible “AI generated” label. A reader cannot establish the presence or absence of the mark merely by looking for a badge, odd spacing, or unusual punctuation.
Does it survive copy and paste?
Anthropic says the watermark is part of the text and will travel when the text is copied and pasted elsewhere. A move from Claude to a document, message, website editor, or plain-text field does not by itself guarantee removal.
This behavior is one reason text watermarking differs from file metadata, which may be discarded when content is separated from the original file.
What happens when text is edited?
Anthropic says the watermark may persist through some editing. It also acknowledges that heavily edited, paraphrased, translated, or mixed text might no longer contain a detectable mark.
There is no published threshold that tells users how much change is “some” or “heavy.” Passage length also matters: Anthropic says very short text may not carry enough information for a reliable signal.
Which Claude models use marking?
Anthropic says Claude models launched in the EU on or after August 2, 2026 support marking at launch. It is also working to add marking support to models released before that date.
For supported models, Anthropic says marking applies wherever Claude is offered worldwide and across Claude products, API access, and listed cloud partners, although some platforms or features may not support every marking type.
What is currently known about detection?
Anthropic says it is working to let users and third parties check for embedded watermarks and provenance metadata. A supported positive result would indicate that content may have been processed by Claude.
The word “may” matters. Claude can proofread, translate, summarize, or transform material that originated elsewhere, and marked content can be altered or combined after processing.
What Anthropic has not publicly disclosed
The public article does not specify how the text signal is encoded, how a detector calculates a result, what confidence threshold it uses, or what exact passage length is required.
Until technical documentation is released, independent implementations cannot be validated against an official specification. This site does not fill those gaps with guesses.
Text watermarking versus file and C2PA provenance
Claude’s text watermark is described as embedded in generated text. For supported files such as SVG, PNG, and JPG, Anthropic instead describes signed provenance metadata based on the C2PA open standard.
Provenance metadata can signal that a file was processed by Claude and whether the signed file has been tampered with. It is separate from the text watermark and may be stripped when a file is converted, re-saved, or captured as a screenshot.
What this site can inspect
This site can inspect ordinary Unicode formatting characters by exact code point. Standard coverage includes exotic spaces U+00A0, U+1680, U+2000–U+200A, U+202F, U+205F, and U+3000; separators U+2028–U+2029; and selected invisible characters including U+00AD, U+200B, U+2060, U+FEFF, and U+E0001–U+E007F.
Broad coverage additionally includes U+034F, U+061C, U+200C–U+200F, U+202A–U+202E, U+2061–U+2064, U+2066–U+206F, and U+FFF9–U+FFFB. These findings are supported formatting signals, not proof of Anthropic’s proprietary watermark or Claude authorship.
How Claude Watermarking Works: FAQ
Is the Claude text watermark a visible label?
No. Anthropic describes it as imperceptible and machine-readable.
Is the watermark stored only in Claude’s interface?
No. Anthropic says it is embedded in the text itself and travels when supported marked text is copied and pasted.
Has Anthropic explained the technical encoding?
Not in its current public documentation. More detailed technical and detection guidance is forthcoming.
Does a detected mark prove Claude wrote everything?
No. Anthropic says detection indicates content may have been processed by Claude and is not fully conclusive about provenance or original authorship.
Are file provenance labels the same as text watermarks?
No. Anthropic describes embedded text watermarks and signed C2PA provenance metadata as complementary but different marking methods.