Anthropic has set out how it plans to mark content generated or processed by Claude, pairing text watermarks with signed provenance metadata for selected files. The company frames the work as part of its commitments under the European Union’s Article 50(2) Code of Practice on Transparency of AI-Generated Content, which it signed as a provider of generative AI models and systems.
The announcement matters because it moves the discussion beyond a broad promise to label AI output. Anthropic describes two technical mechanisms, the types of content they are intended to cover, and several reasons a mark should not be treated as a definitive answer to who created a piece of work. The company says more detailed technical guidance and detection information will follow.
Two signals, aimed at different kinds of output
For text, Anthropic says supported Claude models will embed an imperceptible watermark in the output. The company says the mark does not change the text’s meaning, quality or readability. Because it is woven into the text, it may continue to travel with material that is copied and pasted, and may survive some subsequent editing.
The company says this will be applied at the model level, rather than being limited to a particular Claude interface. That distinction is significant for organisations using Claude through different products or integrations: the intended signal is tied to supported model output, not simply to a label shown in one consumer application.
For supported files, including SVG, PNG and JPG formats, Claude will attach signed provenance metadata. Anthropic says this metadata uses the Coalition for Content Provenance and Authenticity (C2PA) open standard. When the metadata is present, it can indicate that a file was processed by Claude and provide a way to detect whether the signed record has been tampered with.
Provenance is not the same as authorship
Anthropic is unusually explicit that a Claude mark cannot, by itself, settle every question about where content came from. A person may use Claude to proofread, translate, summarise or convert existing material. In those cases, the resulting output can carry a Claude mark even when the underlying ideas, data or original writing came from somewhere else.
That is a useful distinction for publishers, educators, compliance teams and customers reviewing external work. A provenance signal can provide context about how a file or passage was processed. It is not a universal detector of AI authorship, nor is it evidence that Claude created every component of a marked item from scratch.
The support guidance also lists situations where a detection result may be unavailable or unreliable. Text may have been heavily edited, paraphrased, translated or mixed into other writing. Short passages may not contain enough material for a reliable text signal. File metadata can disappear when a document is converted, re-saved or captured in a screenshot. Older models, unsupported features and unsupported file types are further gaps.
Detection tools are still to come
Anthropic says it is working on ways for users and third parties to detect both the embedded text watermarks and the provenance metadata. For now, the company has not published the operational detail needed to assess how detection will work in practice: for example, which models and file types will be supported first, how customers will access checks, the expected performance under common editing workflows, or the timetable for rollout.
Those details will shape the practical value of the announcement. C2PA metadata is designed to support interoperable provenance records, but its usefulness depends on retaining that metadata across the tools used to create, share and publish files. Text watermarking faces a different problem: its value will depend on whether detection remains dependable after normal human revision without creating false confidence about authorship.
An EU transparency commitment takes a product form
The policy context is the EU AI Act’s transparency rules for certain AI-generated content. Article 50 has focused attention on machine-readable signals that can help people understand when they are encountering synthetic material. Anthropic’s statement does not turn every Claude output into visibly labelled content; instead, it describes underlying machine-readable mechanisms and acknowledges their limits.
For organisations building with Claude, the announcement is not a substitute for their own legal or product assessment. Anthropic says customers deploying Claude should independently consider what Article 50 requires for their own services. The company’s stated aim is to help those customers meet relevant transparency obligations, with further technical guidance to come.
That leaves important implementation questions open for Australian organisations as well. The EU commitment is not an Australian legal requirement, but product teams operating across markets may still need a consistent approach to provenance, user disclosure and records of AI-assisted content. Teams should avoid assuming that a Claude mark will cover downstream publishing obligations, especially if their workflow transforms text or strips image metadata.
A measured step toward content transparency
Anthropic’s plan is a material product and policy development because it describes persistent signals built into supported Claude outputs, rather than a temporary interface notice. Yet the company’s own caveats are central to the story. A mark can show that Claude processed content under supported conditions; it cannot reconstruct the whole creation history, identify every contributor or reliably withstand every transformation.
The next useful milestone will be the technical documentation: supported models, availability, detection methods and testing results. Until then, organisations should regard the new marks as one source of provenance context, alongside their own disclosure policies, editorial review and record-keeping practices.