What Anthropic announced
Anthropic signed the EU AI Act Article 50(2) Code of Practice on Transparency of AI-Generated Content and described how it is introducing machine-readable marks in supported Claude output.
The company describes two forms of marking: an imperceptible watermark embedded in generated text and signed provenance metadata for supported files. These are technical signals rather than visible labels placed over an answer.
How Claude text marking works at a high level
Anthropic says the text watermark is woven into output from supported models without changing its meaning, quality, or readability. Because the mark belongs to the text, it can travel when marked text is copied and pasted into another application.
The public documentation does not explain the encoding method. It therefore does not support claims about a particular hidden-character pattern, token rule, or independent removal technique.
Which outputs are affected
Generated text from supported Claude models can contain the embedded watermark. For supported SVG, PNG, and JPG files, Anthropic describes signed provenance metadata based on the C2PA open standard instead of the text-marking method.
Support can vary by model, platform, feature, and output type. A supported model also does not guarantee that every short or substantially changed passage will retain enough information for later detection.
Current rollout status
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 eligible models released before that date.
For models that support marking, Anthropic says the marks apply worldwide wherever Claude is offered, including Claude products, API access, and supported cloud-partner access. The official statement is based on rollout timing rather than a permanent model-name list.
What Anthropic has said about detection
Anthropic says it is working on ways for users and third parties to check for embedded watermarks and provenance metadata. Its current public article says more detailed technical and detection guidance is forthcoming.
A future positive result would indicate that content may have been processed by Claude. It would not, by itself, prove original authorship or establish that every word was generated by Claude.
What remains unknown
Anthropic has not publicly specified how the text signal is encoded, how a detector calculates a result, what confidence threshold it would use, or the exact passage length needed for reliable detection.
It also does not publish a fixed editing threshold. Anthropic notes that heavily edited, paraphrased, translated, mixed, or very short text might not contain a detectable mark.
What Reddit discussions are asking about Claude watermarks
Current Claude watermark discussions on Reddit repeatedly ask whether copy and paste removes the mark, whether Claude Code and API output are covered, whether ordinary editing changes detectability, and whether a public checker already exists.
Anthropic’s documentation answers only part of that discussion: supported marking is applied at the model level, copied text can retain the signal, substantial editing may affect detectability, and public technical detection details are still forthcoming. Claims about a specific encoding or guaranteed removal remain unverified.
What this means for Claude users
Users should treat Claude marking as a developing provenance signal, not as a visible label or conclusive authorship test. Copying text alone does not necessarily remove an embedded mark, while substantial changes can affect detectability.
This site will update its coverage when Anthropic publishes meaningful changes. Its browser tool performs transparent text normalization and does not claim to implement Anthropic’s unpublished detector.