Anthropic has announced plans to add invisible watermarks and signed metadata to content created by Claude AI. The move follows the company’s decision to sign the European Union AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content.
The system is designed to help people identify content that may have been generated or processed by Claude. Claude models launched in the European Union on or after August 2, 2026, will support machine-readable marking from their release date.
Anthropic said the marking system will apply globally, not only in Europe. It will cover supported Claude models used through the Claude website, API platform, Claude Code, Claude Cowork, Claude Tag, and cloud-based services.
The company will use 2 separate marking methods. The first method is an invisible watermark embedded directly into AI-generated text. The watermark will not be visible to readers. It should not alter the meaning, quality, style, or readability of the text.
Anthropic Invisible Watermarks To Outputs
Anthropic said the watermark is built into the model output itself rather than added only by a specific application. This approach ensures the watermark remains visible when users copy and paste Claude-generated text into another document, website, email, or application.
It may also survive certain edits. However, the company warned that extensive rewriting, paraphrasing, translation, or mixing the content with human-written material could weaken or remove the detectable signal.
The second method involves digitally signed provenance metadata for supported files. Anthropic plans to attach this metadata to files such as .svg, .png, and .jpg images when Claude generates or processes them.
The metadata will use the Coalition for Content Provenance and Authenticity (C2PA) open standard. C2PA is an industry-backed framework intended to record information about how digital content was created or modified.
MethodPurposeContentText WatermarkHidden, machine-readable markingAI TextC2PA MetadataVerifies origin and changesSVG, PNG, JPG//
A valid, signed label can indicate that Claude processed a file and help identify whether metadata has been altered. However, metadata protection has technical limits.
It can be removed when files are converted, saved again, edited through unsupported tools, or captured in screenshots. Signed metadata may also be unavailable on some cloud platforms or file-processing features.
Anthropic said text watermarks will also apply when Claude models supported by AWS, Google Cloud, or Microsoft Foundry are accessed.
File provenance metadata may differ between these platforms because cloud providers offer different product features and file-handling capabilities. The company is also developing tools that will allow users and third parties to detect Claude watermarks and provenance information.
A successful detection result would indicate that Claude may have processed content. It would not prove that Claude was the original author, because users can submit human-created text, research, images, or files to Claude for summarizing, editing, translation, or proofreading.
Likewise, the absence of a watermark should not be treated as proof that a human wrote content. Older Claude models may not yet support marking, short text may contain too little material for reliable detection, and heavily edited output may lose its signal.
Anthropic said it is working to add marking support to models released before August 2, 2026, during the transition period allowed under the EU AI Act.
For developers building products with Claude, Anthropic said they must still assess their own Article 50 transparency obligations. The company plans to publish more technical guidance on watermark detection, supported file types, and implementation details as its marking system becomes available.
Strengthen Your SOC by Accelerating Threat Detection & Rapid Investigations. -> Integrate ANY.RUN With Your SOC Now.
The post Anthropic to Add Invisible Watermarks to All Claude AI Outputs appeared first on Cyber Security News.
