Anthropic is introducing machine-readable watermarks in Claude-generated text, allowing supported AI output to carry an invisible signal that can be detected even after the text is copied elsewhere.
The system is part of Anthropic’s implementation of the European Union AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content, which the company has signed as a provider of generative AI models and systems.
Claude models launched in the EU on or after August 2, 2026, will support marking from release, while Anthropic says it is also working to extend the system to older models during the law’s transition period.
The markings will apply worldwide across supported Claude products, including Claude, the Claude Platform API, Claude Code, Claude Cowork, and Claude Tag. Anthropic says text watermarks will also apply when supported models are accessed through AWS, Google Cloud, or Microsoft Foundry.
An invisible watermark embedded in text
For ordinary text, Claude will embed an imperceptible watermark directly into generated output.
Anthropic has not yet disclosed the technical details of the watermarking method, but says the marker does not visibly alter the text or affect its meaning, quality, or readability.
Because the signal is embedded in the text itself rather than stored separately as metadata, it can remain when someone copies and pastes the content into another application and may survive some editing.
The watermark is implemented at the model level, meaning supported Claude models should add it regardless of which Anthropic product or compatible cloud service generated the response.
Anthropic is also developing detection tools that will allow users and third parties to check text for supported Claude marks, with more technical documentation expected later.
However, detecting a watermark will not prove that Claude originally authored the content.
For example, a human-written article could receive a Claude mark after the model has proofread, translated, summarized, or reformatted it. Content could also be modified after Claude processed it.
Likewise, the absence of a detectable mark does not prove that text was written by a human. Heavy editing, paraphrasing, translation, mixing AI text with other material, very short passages, or output from older unsupported models could remove or weaken the signal.
Generated files get signed provenance metadata
For supported file types such as PNG, JPG, and SVG, Anthropic plans to attach digitally signed provenance metadata based on the Coalition for Content Provenance and Authenticity (C2PA) standard.
This metadata can indicate that a file was processed by Claude and help reveal whether it was altered afterward.
Unlike the embedded text watermark, however, file metadata can be stripped through format conversion, re-saving, screenshots, or software that does not preserve it. Anthropic also notes that signed provenance metadata may not be available through every platform or integration.
A detected mark can indicate that content passed through Claude, but it cannot establish who originally wrote it or how much AI involvement there was.
Anthropic says it will publish additional technical guidance explaining how third parties can detect both the embedded text watermarks and signed provenance metadata.
Leave a Reply