4 minNews
Anthropic Moves Claude Toward Machine-Readable Provenance
Claude’s supported text outputs will carry invisible marks and supported files will carry signed provenance. The rollout is tied to new EU transparency rules.
Anthropic is preparing Claude to leave a machine-readable record of its involvement in generated content, a change driven in part by new European transparency rules. Supported Claude models will embed an imperceptible watermark in text, while supported generated files will carry digitally signed provenance metadata.
The change should not be described as a blanket rollout that began for every Claude model on August 2. That date marks the start of the European Union’s Article 50 transparency obligations. Anthropic says Claude models released on or after August 2, 2026 support marking from launch. Previously released models are still being brought into the system, and the EU provides a limited transition period through December 2 for the machine-readable marking requirement on older systems.
Article 50 requires providers of generative AI systems to make synthetic audio, images, video and text machine-readable and detectable as artificially generated or manipulated, subject to technical feasibility and the law’s scope. The policy reflects a broader effort to give platforms and the public a technical way to distinguish synthetic content without relying solely on a creator’s voluntary disclosure.
Anthropic says its text watermark is inserted at the model level and does not change the meaning, quality or readability of the output. The mark is designed to travel with copied text and may survive some editing. That makes it fundamentally different from a label displayed by a website, which disappears as soon as the words are pasted somewhere else.
The company has not published the detailed method behind the watermark. It has also not yet released the full set of tools third parties will use to detect it. As a result, claims that the system consists of a simple set of invisible characters go beyond what Anthropic has publicly described.
Generated files use a separate approach. Where supported, Anthropic will attach digitally signed provenance information. For supported media, it uses the C2PA standard, which can bind cryptographically verifiable claims about an asset’s origin and history to the file.
That provenance record has limits. Metadata can be lost or stripped when a file is reprocessed, and a valid provenance record does not certify the truth of the content. It can indicate that a file came through a particular tool or workflow, but it cannot establish whether a factual statement inside the file is correct.
The text watermark is also not infallible. Substantial rewriting, translation or combining Claude output with other text may reduce detectability. The absence of a mark cannot therefore be treated as proof that a human wrote the material. A positive signal also does not show how much of the final work came from Claude; the model may have been used for editing rather than original drafting.
Anthropic says the marking system will apply globally to supported Claude models, including across the API and supported cloud delivery. The choice means a rule triggered by EU legislation will influence the behavior of Claude beyond Europe.
For institutions, the practical challenge comes next. Schools, publishers, employers and platforms will need standards for interpreting a detected watermark, including cases where AI was used only in part of a workflow. Until Anthropic releases its detection tooling and the system is tested under ordinary editing conditions, the new mark should be understood as evidence about provenance—not a final verdict on authorship.
