Anthropic has announced plans to introduce invisible text watermarks and cryptographically signed metadata for content generated or processed by its Claude AI models, expanding transparency controls intended to help users identify AI involvement in digital content.
The initiative follows Anthropic’s decision to sign the European Union AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content.
Under the plan, Claude models released in the European Union on or after August 2, 2026, will include machine-readable content-marking from the date they become available.
Anthropic Adds Invisible Watermarks and C2PA Metadata
Although prompted by European transparency requirements, Anthropic said the marking capabilities will be deployed globally.
The controls are expected to apply across supported Claude models accessed through the Claude website, API platform, Claude Code, Claude Cowork, Claude Tag, and supported cloud-hosted services.
Anthropic’s approach consists of two separate technical mechanisms: invisible watermarking for text and C2PA-based provenance metadata for files.
The first mechanism embeds an invisible, machine-readable watermark directly into Claude-generated text. Unlike a visible disclosure label, the watermark is not intended to affect readability, meaning, style, or output quality.
It is designed to remain undetectable to ordinary readers while allowing authorized detection tools to identify that Claude may have generated or processed the material.
Anthropic said the watermark is generated within the model output rather than attached by a particular application or interface. This model-level implementation means the signal may persist when users copy Claude-generated content into documents, websites, emails, messaging platforms, or other applications.
The company cautioned, however, that the watermark is not immutable. Extensive rewriting, translation, paraphrasing, or blending Claude output with human-authored content may reduce or remove the detectable signal.
Short passages may also contain insufficient material for reliable watermark detection. The second method involves digitally signed provenance metadata attached to supported files, including SVG, PNG, and JPG images generated or processed through Claude.
Anthropic plans to use the Coalition for Content Provenance and Authenticity (C2PA) standard, an industry-backed framework for recording information about a digital asset’s origin and modification history.
A valid signed provenance label can indicate that Claude handled a file and may help reveal whether associated metadata was modified after creation.
| Method | Purpose | Covered Content |
|---|---|---|
| Invisible watermark | Hidden, machine-readable AI marking | Text |
| C2PA signed metadata | Origin and modification provenance | SVG, PNG, JPG files |
C2PA metadata also has practical limitations. It can be stripped when files are converted, re-saved, edited with unsupported tools, or captured through screenshots. Metadata availability may further vary across cloud services and file-processing workflows.
Text watermarking will also apply when supported Claude models are accessed through AWS, Google Cloud, and Microsoft Foundry. File-provenance support may differ across these environments because cloud providers expose different features and apply distinct file-handling processes.
Anthropic is developing tools for users and third parties to detect Claude watermarks and provenance information. However, a successful detection should not be interpreted as conclusive proof that Claude originally authored the content.
Users may submit human-written text, images, research, or files to Claude for editing, translation, summarization, or proofreading. Similarly, the absence of a watermark does not establish human authorship. Earlier Claude models may not support marking, and signals can be degraded through extensive modification.
Anthropic said it is working to add marking support to models released before August 2, 2026, during the EU AI Act transition period. Developers integrating Claude must still assess and meet their own Article 50 transparency obligations.
Detect, investigate, and respond faster with in-browser data inspection from ANY.RUN. Gain complete phishing visibility to strengthen your SOC and reduce MTTR
