What happens to a text that Claude writes? Since August 2, it carries a marking that no human can read, but every machine searching for it can find. Anthropic has become the first major commercial AI company to embed a technical watermark directly into the model layer, not as a post-hoc stamp on text. This means: whoever copies the text and pastes it elsewhere takes the mark along.
What the Watermark Is and How It Works
The watermark sits within the text itself, not in a footnote or metadata set. Anthropic describes it as an imperceptible signal, a non-perceptible signal component that remains completely invisible during normal reading but can be detected by recognition software. TechCrunch quotes the company: "The watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere."
For image files, Anthropic relies on the C2PA standard (Coalition for Content Provenance and Authenticity), an open industry standard already used by Adobe, Microsoft, and Google for image signatures. This standard embeds cryptographically secured provenance information in metadata. Who created it, when, and with what tool becomes machine-verifiable.
The watermark applies to all products: Claude App, Claude Platform API, Claude Code, Claude Cowork, and Claude Tag, plus Claude models delivered via Amazon Web Services, Google Cloud, and Microsoft Foundry. All new models released after August 2, 2026 have the feature automatically. Older models are undergoing retrofitting, which Anthropic has committed to for the EU transition period.
Why Now and Why Globally
The trigger is Article 50 of the EU AI Act. The so-called Transparency Code took effect on August 2, 2026, and mandates that AI-generated content must be machine-readable labeled. Specifically, this applies to all providers and operators of AI systems that produce or substantially modify text, images, video, or audio.
Anthropic chose global implementation, not limited to European users. Euronews quotes the company: the labeling applies to all services worldwide, regardless of the user's location. This is a business decision, not a legal obligation outside the EU, but it substantially simplifies technical architecture because geographic differentiation isn't necessary.
Anthropic isn't alone. Google, Meta, Microsoft, and OpenAI have also announced or already implemented watermarking solutions. Music platform Suno and writing platform Substack have made similar announcements. The EU AI Act thus creates an industry standard through regulatory pressure that extends well beyond Europe.
What Users Feel and What They Don't
In everyday use, little changes for Claude users initially. Generated text looks identical. Quality should not be affected, according to Anthropic; the watermarking technique is lossless with respect to readability. What changes is traceability.
Anthropic has announced it will publish technical details about watermark detection so third parties, including newsrooms, platform operators, and educational institutions, can develop their own verification tools. When exactly those details appear, the company hasn't specified.
Reaction from parts of the user base is critical. Forbes captures sentiment on social networks under the title "The internet isn't happy." Central objections include: it's unclear how stable the watermark remains under heavy text editing. Does translation into another language remove it? How does it hold up with intensive rephrasing? Anthropic hasn't answered these questions yet.
The Open Question of Detectability
EU law prescribes no specific labeling form. What's required is clear and unambiguous machine-readable information. Whether an invisible text watermark meets requirements, and thus is sufficiently robust and reliably detectable, will be clarified first by national regulators and likely by initial court rulings. The Register notes that Anthropic hasn't answered "how much editing is needed to remove the watermark."
For news organizations, platform operators, and schools, this is the decisive practical question: if a student lightly rewrites a Claude text, does the detection system still show AI involvement? If a journalist revises AI drafts, does the result count as AI-generated? Germany's Digital Economy Association (BVDW) is developing voluntary industry standards, though these won't be legally binding. By fall 2026, the first test cases will show how robust the technology truly is.