Anthropic adds invisible watermarks to output from new Claude models
Anthropic will add machine-readable marks to output from Claude models launched on or after August 2, 2026.
Anthropic will embed invisible watermarks in text generated by new Claude models and attach signed provenance metadata to supported files, implementing commitments under the European Union’s new transparency requirements for AI-generated content.
The policy covers Claude models launched on or after August 2, 2026. Although prompted by Article 50 of the EU AI Act, marking will apply wherever supported Claude models are offered worldwide, rather than only to output produced in the EU.
Anthropic has signed the Article 50(2) Code of Practice on Transparency of AI-Generated Content as a provider of both generative AI models and generative AI systems. The company’s plan uses two forms of machine-readable marking, one embedded directly in generated text and another attached to files.
Outputs will be marked across Anthropic’s products, including the Claude Platform API, Claude, Claude Code, Claude Cowork and Claude Tag. Embedded watermarks will also apply when supported models are accessed through AWS, Google Cloud or Microsoft Foundry.
Support is not identical across every service. Anthropic says signed provenance metadata may be unavailable through some cloud platforms, products, features or file types.
Invisible watermarks follow copied Claude text
When a supported model generates text, Claude will weave an imperceptible watermark into the output. Anthropic says the mark does not alter its meaning, quality or readability.
Because marking happens at model level, the watermark remains part of the text regardless of which Claude product produces it. It will travel when the text is copied and pasted and may survive some subsequent editing.
Supported files, including SVG, PNG and JPG formats, will instead receive digitally signed provenance metadata based on the Coalition for Content Provenance and Authenticity’s C2PA open standard.
A valid label signals that Claude processed the file and can indicate whether it has been altered. However, Anthropic stops short of presenting that label as proof that Claude originated everything within the file.
A Claude mark is a signal, not proof of authorship
As people also use Claude to proofread, translate, summarize and convert existing materialoutput may carry a Claude mark even when its central ideas, text or data came from another source.
Marked content can also be edited, excerpted or combined with other material after leaving Claude. A detected mark consequently indicates that Claude may have processed the content, rather than establishing its complete history.
The reverse is equally limited. Failing to detect a mark does not demonstrate that content was written or produced without AI.
Anthropic says a mark may be lost when text is heavily edited, paraphrased, translated or mixed with other writing. Very short passages may not contain enough material for reliable detection, while provenance metadata can disappear when a file is converted, resaved or captured in a screenshot.
Content produced by older Claude models or through unsupported platforms, features and file types may also have no detectable mark.
Anthropic is working to add marking support to models released before August 2 under the legislation’s transition arrangements. It is also developing detection methods for users and third parties, although technical documentation and a release timetable have not yet been published.
Developers deploying Claude within their own products remain responsible for assessing how Article 50 applies to their services. Anthropic says further technical guidance on its marking and detection approach will be released as it becomes available.
