Anthropic Claude watermarks will label new EU model output
Anthropic plans invisible text marks and signed file metadata for new Claude models, but says detection cannot prove authorship.
By Dominic Okoye · Staff Writer
· 3 min read
Anthropic says new Claude models released in the EU from August 2, 2026 will add machine-readable labels to their output under its commitments to the EU AI Act’s transparency code. The Anthropic Claude watermarks program will place imperceptible marks in generated text and digitally signed metadata on supported files, though the company says neither a positive nor negative result can definitively settle whether content originated with AI.
The company has signed the EU AI Act Article 50(2) Code of Practice on Transparency of AI-Generated Content as both a generative-model provider and a generative-system provider, according to its support documentation. Anthropic says models released before the August 2 date are still being updated during a transition period.
How will Anthropic Claude watermarks work?
Anthropic describes two forms of marking. Text produced by supported Claude models will contain an invisible watermark embedded in the text. The company says the mark does not alter the meaning, quality or readability of an answer, travels when text is copied and pasted, and may survive some editing.
For supported files, including .svg, .png and .jpg formats, Anthropic says it will attach digitally signed provenance metadata using the Coalition for Content Provenance and Authenticity, or C2PA, open standard. Anthropic says the signed label can signal that a file was processed by Claude and allow detection of tampering.
Coverage is intended to extend across the Claude Platform API, Claude, Claude Code, Claude Cowork and Claude Tag. Text watermarking also applies when supported models are accessed through AWS, Google Cloud and Microsoft Foundry, Anthropic says. The company says the program applies worldwide wherever supported Claude models are available, although specific platforms and features may not support every marking method.
Can a Claude mark prove who made the content?
No. Anthropic says a detected mark is a signal that material may have been processed by Claude, rather than conclusive evidence that Claude was the original author. A user might send human-created text or data to Claude for proofreading, translation, summarization or file conversion, and the resulting output could still carry a mark.
An absent mark is similarly not proof that material was not AI-generated or processed by Claude. Anthropic says a mark may not be detectable in output from earlier models, very short passages, or text that has been heavily edited, paraphrased, translated or combined with other writing. File metadata may also be lost after format conversion, re-saving or screenshots.
That leaves the marks as a provenance input for publishers, platforms and enterprise buyers, not a general-purpose AI-content verdict. Anthropic has said it will provide users and third parties with ways to detect the marks, but has not yet published the promised technical documentation or detection details.
For teams building products on Claude, the company says they must separately assess what the EU’s Article 50 requires of their own services. Anthropic’s announced labeling can support a transparency workflow, but its stated limitations mean downstream operators cannot treat it as a clean authorship record.
This story draws on original reporting from The Register.
